Skip to main content
Glama

Zebbern Kali MCP Server

Give an AI agent a full Kali Linux toolkit without installing one. The MCP client runs on your machine as a small Python package; every tool executes inside a Kali container that the client reaches over HTTP. Nmap, sqlmap, nuclei, impacket, Metasploit, pivoting, VPN and callback capture are all exposed as MCP tools.

License Python MCP tools Base image

Quick start · Essentials (jobs, files, VPN, targets) · Configuration · Troubleshooting · Tool modules · Image contents · Security · Development


Quick start

1. Start the Kali backend

docker run -d --name zebbern-kali --restart unless-stopped \
  -p 127.0.0.1:5000:5000 \
  --cap-add=NET_ADMIN --cap-add=NET_RAW --device=/dev/net/tun \
  ghcr.io/zebbern/zebbern-kali-mcp:latest

NET_ADMIN, NET_RAW and /dev/net/tun are what OpenVPN and WireGuard need; without them vpn_connect fails whatever the config says. Add any of these flags for full parity with the Compose file:

  • -p 127.0.0.1:1080:1080 — published SOCKS proxy

  • --sysctl net.ipv4.ip_forward=1 — pivoting and routing

  • --shm-size=1g — headless Chrome; without it gowitness writes no screenshot

  • -v zebbern-kali-tmp:/app/tmp — keep background-job logs across recreates

  • -v "$(pwd)/vpn:/vpn:ro" — VPN configs from the host

From a source checkout, Compose does all of the above:

git clone https://github.com/zebbern/zebbern-kali-mcp.git
cd zebbern-kali-mcp
docker compose up -d           # add --build to build the image locally

2. Point your MCP client at it

{
  "servers": {
    "kali-tools": {
      "command": "uvx",
      "args": ["zebbern-kali-mcp"]
    }
  }
}

uvx downloads the client from PyPI on first run; nothing to install by hand. VS Code reads .vscode/mcp.json (or your global MCP config) with the servers key above, Claude Desktop and ~/.claude.json use mcpServers with the same object, and CLI flags go in args, e.g. ["zebbern-kali-mcp", "--profile", "web"].

Pin the version — ["zebbern-kali-mcp@<version>"] — to make upgrades explicit. Unpinned, uvx reuses whatever environment it already cached for that command, so a restart does not necessarily pick up a newer wheel; see Troubleshooting.

Restart the client. That is the whole install.

3. Ask

"Scan 10.10.10.5 with nmap" · "Run nuclei against example.com" · "Connect to the HTB VPN and start recon" · "Enumerate AD with bloodhound against dc01.corp.local" · "Start a callback listener on port 8080"

Call health to confirm the client reached the backend.


Related MCP server: Hercules MCP

Essentials

Four things that are not obvious from the tool list.

Long-running tools hand back a job_id

An MCP host abandons a synchronous tool call after roughly a minute, so anything slower runs as a background job on the backend. Seventeen tools do that for you: the fourteen heavy scanners (tools_nmap, tools_nikto, tools_gobuster, tools_wpscan, tools_sqlmap, tools_hydra, tools_masscan, tools_katana, tools_amass, tools_arjun, tools_fierce, tools_enum4linux, tools_gowitness, tools_john) plus api_nuclei_scan, api_ffuf_fuzz and zebbern_exec. The job starts first, the client waits inline for ZKM_INLINE_WAIT_SECONDS (default 50), and you get back either the finished result or a job_id. A wait that is too short costs a poll, never the scan. Any other subprocess-backed tool takes background=true explicitly.

Call

Purpose

job_status(job_id)

state, exit code, timing

job_output(job_id)

output so far (long-polls briefly)

send_input(job_id, text)

write to the process's stdin

job_cancel(job_id)

kill the job's process group

job_list()

every job the backend still tracks, newest first — use it when the job_id is gone

Output is never capped. The response window is bounded, but every byte of stdout and stderr is also teed to $JOB_OUTPUT_DIR/<job_id>.log (/app/tmp/jobs in the image), so output_truncated does not mean lost output. output_logged: false does: the log directory was not writable and only the bounded window exists. Those logs are never rotated or pruned — clean up with rm /app/tmp/jobs/*.log or by recreating the container.

Check timed_out, not success, to know whether a command finished: a truncated scan reports success: true and timed_out: true on purpose, because its partial output is worth keeping.

exec_stream streams output but registers no job, so it cannot be cancelled — on disconnect the subprocess runs to its full timeout, untracked. Use zebbern_exec(background=true) for anything you may need to abort. Job state, listeners and sessions live in memory, so a backend restart drops them and the *_status tools then answer empty rather than erroring.

Getting files in and out

kali_upload(content, remote_path) writes a file into the container — a config, wordlist, script or target list, anything whose quoting or newlines a shell command line would mangle. content is always base64, so base64-encode the UTF-8 bytes of plain text first; kali_download(remote_path) reads a file back out as base64.

target_upload_file / target_download_file move files to and from a remote target instead, and ssh_session_* / reverse_shell_* transfer over an existing session.

VPN configs

Put .conf (WireGuard) or .ovpn (OpenVPN) files in the host directory mounted read-only at /vpn (VPN_DIR, default ./vpn), then call vpn_connect(config_path="/vpn/client.ovpn"). The type is auto-detected from the file contents; override with vpn_type="wireguard" or vpn_type="openvpn".

No mount is required: the entrypoint creates /vpn on every boot, so docker cp client.ovpn zebbern-kali:/vpn/client.ovpn works on a container started without one.

Reaching your targets

Inside the container 127.0.0.1 is the container's own loopback — scanning it finds an empty container, which looks exactly like a broken MCP server. The host, and anything published on it, is host.docker.internal. Services on another Compose network that are not host-published need the Kali container attached to that network.


Configuration

Environment variables

Variable

Default

Description

KALI_API_URL

http://127.0.0.1:5000

MCP client: URL of the Kali Flask server

MCP_TOOL_PROFILE

auto

auto, core, recon, web, ad, ctf, trim, full

MCP_EXCLUDE_MODULES

(empty)

Modules to drop from the selected profile, e.g. callback_catcher,output_parser

ZKM_INLINE_WAIT_SECONDS

50

How long an auto-promoting tool waits inline before returning a job_id

API_PORT

5000

Flask server port

API_BIND_ADDRESS

127.0.0.1

Host address for the Compose API port publication

API_LISTEN_HOST

0.0.0.0

API listener inside bridge mode; host-network mode defaults to 127.0.0.1

KALI_API_TOKEN

—

Optional shared API token; enforced on /api/* only when set

DEBUG_MODE

0

Enable debug logging

REQUIRED_TOOLS

—

Comma-separated binaries that must exist for /ready to report ready

HTB_ROUTES

—

Comma-separated CIDRs to route, e.g. 10.129.0.0/16,10.10.0.0/16

EXTRA_HOSTS

—

Comma-separated hostname:ip pairs added to /etc/hosts

VPN_DIR

./vpn

Host directory mounted read-only at /vpn. Optional — /vpn is created at boot regardless, so docker cp works without a mount

SOCKS_BIND_ADDRESS

127.0.0.1

Host address for the Compose SOCKS port publication

SOCKS_PORT

1080

Published host SOCKS port

SOCKS_LISTEN_HOST

0.0.0.0

SOCKS listener inside bridge mode; host-network mode defaults to 127.0.0.1

JOB_OUTPUT_DIR

/app/tmp/jobs in the image

Where full job logs are teed; falls back to the OS temp dir from source

JOB_MAX_COUNT

256

Maximum retained background jobs

JOB_OUTPUT_MAX_LINES / _MAX_CHARS / _MAX_LINE_CHARS

2000 / 2097152 / 4096

Bounds on the retained response window per job: output events, characters, characters per event

JOB_INPUT_MAX_BYTES / JOB_INPUT_QUEUE_SIZE

65536 / 16

Maximum input bytes per job request, and queued input requests per job

JOB_OUTPUT_MAX_WAIT

30

Maximum long-poll wait for job output

CTF_MAX_DOWNLOAD_BYTES

104857600

Maximum CTF file download size; a call may request less

INCLUDE_METASPLOIT

true

Build argument: false builds the lean variant

INCLUDE_CADO_NFS

true

Build argument: false builds without CADO-NFS only

Client flags mirror the first few: --server, --profile, --api-token, --exclude-module, --timeout, --debug.

--timeout (default 90000, i.e. 25h) is the HTTP read timeout for a synchronous call: a backstop for a wedged backend, not a scan budget. It must always outlive the backend's own per-tool budget (the longest is 86400s, for hydra and john). Set it lower and the client gives up before the backend can answer — the partial output of a timed-out scan is destroyed and the scan keeps running server-side, orphaned. The connect timeout stays 10s regardless.

Tool profiles

The default auto profile starts from the complete tool set and, given a valid capability manifest (schema version 1), omits only the public tools the backend reports as unavailable — so the backend owns that list instead of the client keeping a parallel copy. On a lean image that is 7 of the 135 tools: the five msf_session_* tools plus payload_generate and payload_templates.

Discovery fails open: unknown, malformed, older or unreachable capability data keeps the complete set, and a manifest that would hide more than half the surface is ignored as a likely backend regression. Core tools — command execution, file operations, host management, output parsing — are never omitted even if a manifest marks them unavailable. The failure directions are not symmetric: a tool that is present but broken fails once and the agent adapts, while a wrongly hidden tool is invisible for the life of the process, because discovery is a startup snapshot. Restart the client to refresh it.

Pick a narrower profile, with --profile web or MCP_TOOL_PROFILE=web, only when a shorter list helps the agent choose tools more reliably:

Profile

Contents

core

Command execution, files, hosts, output parsing

recon

Core plus scanners, fingerprinting, exploit suggestions

web

Core plus web/API testing and callback capture (67 tools)

ad

Core plus AD, pivoting, SSH, shells, payloads, VPN

ctf

Core plus scanners, CTF platforms, payloads, shells, VPN, callbacks

trim

All modules except callback_catcher and output_parser (125 tools)

full

All 17 modules, 135 tools; operator override that ignores discovery

An invalid profile name fails at startup. trim drops the two modules that duplicate what most MCP hosts already provide: callback_catcher (9 tools, overlapping hosted webhook/interactsh services) and output_parser (1 tool, duplicating the agent's own stdout parsing). Prefer full when the host has no webhook capability of its own, or when the engagement runs on an isolated network with no egress — the built-in listener is the only one that works there.

--exclude-module / MCP_EXCLUDE_MODULES subtracts modules from whichever profile is selected, so you can tune the surface without a new profile:

zebbern-kali-mcp --profile web --exclude-module callback_catcher               # 58 tools
zebbern-kali-mcp --profile full --exclude-module callback_catcher,output_parser # 125, same as trim

Names are case-insensitive and whitespace-tolerant; an unknown name fails at startup listing every valid one. Exclusion composes with auto, applying after capability discovery.

Compose, image variants and networking

docker compose up -d                                                    # bridge networking, port-mapped
docker compose -f docker-compose.yml -f docker-compose.host.yml up -d   # host networking
INCLUDE_METASPLOIT=false docker compose build                           # lean variant
INCLUDE_CADO_NFS=false docker compose build                             # faster dev build, CADO-NFS only removed

Build defaults are INCLUDE_METASPLOIT=true and INCLUDE_CADO_NFS=true; both qualified variants (full and lean) include CADO-NFS. With INCLUDE_CADO_NFS=true, a failure to fetch, build or verify the CADO-NFS executable stops the image build.

Compose grants NET_RAW and NET_ADMIN, provides /dev/net/tun, and publishes the API and SOCKS ports on loopback; in host-network mode both services listen on loopback. Set the matching bind or listen variable when another host must connect.

Host networking is qualified on native Linux Docker Engine, which keeps direct host-network semantics, and on the current Windows Docker Desktop 4.84 setup. Docker Desktop needs version 4.34 or later, host networking enabled in Settings > Resources > Network, and a restart before the overlay works; there it supports TCP and UDP only (layer 4), does not work with Enhanced Container Isolation, supports Linux containers only, and cannot bind a specific host-interface IP.

The Kali base image, Go modules, Git sources and standalone downloads are pinned. Kali rolling APT packages and Python transitive dependency resolution are not bit-identical snapshots. linux/amd64 is the only qualified architecture.

For a remote backend, set the same KALI_API_TOKEN in the backend and the client environments, or pass --api-token; direct REST clients send it in the X-API-Key header. Health endpoints stay unauthenticated so Docker and orchestrator checks keep working.


Troubleshooting

"cannot reach the Kali API server" — the backend is not running or KALI_API_URL points elsewhere. The error text includes the full docker run line from Quick start.

A tool is missing, or a documented argument is ignored — suspect a stale client before a bug. Call health: it returns the backend's version plus the client's own client_version, version_match, and a version_note when they disagree. An unpinned uvx zebbern-kali-mcp reuses whatever environment uv already cached for that command, so restarting the MCP server can re-run an older wheel, and --refresh updates the cache without necessarily changing what runs — pin uvx zebbern-kali-mcp@<version> instead. If version_match stays false after a correct pin, an older server process is probably still holding the connection, because hosts start new servers without stopping the old ones. The version in the argv is the version running:

Get-CimInstance Win32_Process | ? { $_.CommandLine -like '*zebbern*' } |
  Select ProcessId, CreationDate, CommandLine

vpn_connect fails whatever the config says — the container is missing --cap-add=NET_ADMIN, --cap-add=NET_RAW or --device=/dev/net/tun.

A scan of 127.0.0.1 finds nothing — that is the container's loopback. Use host.docker.internal.

A long scan or a listener disappeared — jobs, reverse-shell listeners, SSH sessions and MSF sessions live in memory. A backend restart (including docker compose up -d --force-recreate) drops them, and job_list() then answers {"jobs": [], "count": 0}, which means "this backend has run nothing", not "nothing is running". Pivot tunnels are the exception: they persist to state.json and reload as status="stopped".

gowitness reports success but writes no screenshot — headless Chrome needs more shared memory than Docker's 64MB default. Run with --shm-size=1g (Compose already does).

The local image is stale — :latest moves on every backend change. Pull and recreate, then compare the local digest against the remote one:

docker compose pull && docker compose up -d --force-recreate
docker buildx imagetools inspect ghcr.io/zebbern/zebbern-kali-mcp:latest
docker image inspect ghcr.io/zebbern/zebbern-kali-mcp:latest --format '{{index .RepoDigests 0}}'

MCP tool modules

135 tools across 17 client modules in mcp_tools/, each backed by a Flask blueprint in zebbern-kali/api/blueprints/ and core logic in zebbern-kali/core/.

Module

What it covers

kali_tools

Nmap, Nikto, Gobuster, Dirb, WPScan, SQLMap, Hydra, John, enum4linux, Subfinder, httpx, Arjun, Fierce, ssh-audit, gowitness, and more

ad_tools

Active Directory — netexec, BloodHound, impacket, certipy, bloodyAD, Kerberoasting, Pass-the-Hash, LDAP

command_exec

Arbitrary command execution, streaming execution, background jobs, health

ssh_manager

SSH session lifecycle — connect, execute, transfer, disconnect

reverse_shell

Reverse shell listeners, payloads and session management

metasploit

Metasploit Framework — persistent console sessions, module execution

network_pivot

Chisel, Ligolo-ng, SSH tunnels, socat, ProxyChains, SOCKS proxy

vpn

WireGuard and OpenVPN with automatic SOCKS5 proxy

api_security

GraphQL introspection and fuzzing, JWT analysis and cracking, nuclei, ffuf, rate-limit and auth-bypass tests

web_fingerprinter

Technology, header and WAF fingerprinting

exploit_suggester

searchsploit lookups and exploit suggestions from scan results

payload_generator

msfvenom payloads, one-liners, templates, payload hosting

file_operations

kali_upload / kali_download for the container, plus target transfers

callback_catcher

Built-in HTTP + DNS callback listener for isolated networks

ctf_platform

CTFd / rCTF API — challenges, flags, scoreboard, downloads

hosts_management

/etc/hosts inside the container

output_parser

Structured parsing of tool output for AI consumption

zebbern_exec and exec_stream accept any shell command, ssh, scp, rsync, netcat and telnet included — the contract accepts them, it does not promise every binary is bundled in every image variant. The dedicated SSH, pivot and payload managers are conveniences, not restrictions. Nothing is masked in command output or logs.

How it works

HOST (Windows/Linux/macOS)                 DOCKER (kalilinux/kali-rolling)
AI agent → MCP tool (mcp_tools/*)   HTTP   Flask API → api/blueprints/* → core/*
         → KaliToolsClient  ──── POST /api/* :5000 ────→  nmap, sqlmap, msf, …
                            ←──── JSON response ───────

entrypoint.sh sets up routes, /etc/hosts, /vpn, TUN and IP forwarding before the Flask server starts. The client stays a lightweight PyPI package (uvx zebbern-kali-mcp), so the heavy tooling lives only in Docker and never on your host.


What is in the image

Core tool failures stop the build. Tools marked optional fall back to a warning; check /ready and the relevant tool-status endpoint for runtime availability.

Scanning — nmap (service/version detection, NSE), masscan, sslscan, ssh-audit, nikto, gobuster, dirb, wpscan, sqlmap, ffuf, nuclei, katana (v1.1.0 pre-built binary), amass, gowitness, arjun (parameter discovery), net-snmp clients (snmpwalk, snmpget, snmpbulkwalk — the only SNMP primitive here, and no wrapper covers it).

Subdomains, DNS and URLs — subfinder, httpx, assetfinder, waybackurls, waymore, amass, massdns, fierce, mapcidr, subzy (takeover checks).

Passwords — hydra, john, hashcat (GPU-accelerated).

Active Directory — netexec (primary SMB/LDAP/WinRM tool; crackmapexec is deprecated), impacket 0.13.0 (pinned for stable behaviour across rebuilds; ~50 scripts symlinked as impacket-*), bloodhound.py, bloodyAD, certipy-ad (ADCS), responder, evil-winrm, krbrelayx, gMSADumper, PetitPotam, coercer, dementor, winrmexec, pywhisker, ldapdomaindump, man-spider 2.0.0 (manspider — crawls share file content by keyword, regex or extension, where ad_smb_enum only lists shares and tests access; it prints nothing to stdout, logging to ~/.manspider/logs/, so pass -l /app/tmp/manspider-loot to keep loot on the kali-tmp volume rather than the container layer).

Exploitation — metasploit-framework, commix, ghauri, dalfox, byp4xx (403 bypass), git-dumper 1.0.9 (reconstructs a repository from an exposed .git directory, which fingerprint_url only reports the existence of), exploitdb/searchsploit.

JavaScript analysis — getJS, jsluice, xnLinkFinder, SecretFinder, TruffleHog, js-beautify, webcrack (npm), ParamSpider.

API testing and proxies — jwt_tool (/opt/jwt_tool/), graphw00f (GraphQL engine fingerprinting), clairvoyance (schema introspection), mitmproxy (mitmdump), OWASP ZAP (zaproxy), Caido CLI (optional; readiness key caido-cli).

Forensics and CTF — binwalk, steghide, stegseek, zsteg, exiftool, foremost, volatility3, sleuthkit (mmls, fls, icat, blkcat), gdb, radare2, imagemagick, tesseract-ocr.

Binary, crypto and math (Python) — angr, pwntools, pycryptodome, gmpy2, z3-solver, sympy, numpy, scipy, RsaCtfTool (/opt/RsaCtfTool/), cado-nfs (/opt/cado-nfs/, factorization of large keys). SageMath is not bundled in the current Kali rolling image.

Networking and pivoting — scapy, tcpdump, socat, netcat, proxychains4, openvpn, wireguard-tools, chisel (Go binary plus a Windows .exe in /opt/windows-tools/), ligolo-ng v0.7.5 (proxy plus Linux and Windows agents in /opt/ligolo-ng/), cloudflared (optional; readiness key cloudflared), ngrok.

Privilege escalation — LinPEAS (/opt/privesc-tools/linpeas.sh), WinPEAS (x64, x86, .bat, in /opt/privesc-tools/), Mimikatz (/opt/windows-tools/mimikatz/), RunasCs.exe (/opt/windows-tools/).

Cloud, media and containers — awscli, boto3 (importable from the system interpreter: zebbern_exec python3 -c 'import boto3'), ffmpeg, sox (all format plugins), podman (needs --privileged at runtime), Playwright with Chromium for SPA testing, screenshots and JS-rendered pages.

Wordlists — rockyou.txt (decompressed), SecLists, and compatibility symlinks at /usr/share/wordlists/dirb/.

Python dependencies for the container are in requirements.txt; the Dockerfile installs the rest via pip and APT. The image sets NO_COLOR=1, TERM=dumb, FORCE_COLOR=0, CI=true and PWNLIB_NOTERM=1 so tools emit clean, parseable text instead of banners, colours, progress bars and interactive prompts.

Callback catcher

A built-in HTTP + DNS listener for isolated networks where webhook.site or interactsh cannot reach your targets, managed through the callback_catcher module. Defaults are TCP 8888 and UDP 5353. A target can reach them directly through a VPN interface inside the container or with Linux host networking. In bridge mode, publish the callback ports on an address the target can reach (for example 8888:8888/tcp and 5353:5353/udp) — they are not published by default.


Security

This server intentionally provides unrestricted command execution and powerful penetration-testing tools.

The default binds the API and SOCKS ports to 127.0.0.1. Remote use is supported: choose an explicit bind address, set KALI_API_TOKEN, and put TLS or a trusted private network in front when traffic crosses an untrusted one. The container runs as root because several networking and assessment features require it. Use this only on systems you are authorised to test.


Project layout

zebbern-kali-mcp/
├── mcp_server.py           # MCP client entrypoint (FastMCP); ships in the wheel
├── mcp_tools/              # MCP CLIENT, runs on the host — _client.py is the HTTP
│                           #   transport, _autopromote.py the job promotion, then
│                           #   one module per tool category (17)
├── zebbern-kali/           # FLASK SERVER, runs in Docker — kali_server.py,
│                           #   api/routes.py, api/blueprints/ (one per module),
│                           #   core/ (command executor, job manager, tool logic)
├── Dockerfile              # Multi-layer Kali image build
├── docker-compose.yml      # Bridge-mode deployment (+ docker-compose.host.yml)
├── entrypoint.sh           # Container init: routes, hosts, /vpn, TUN, forwarding
├── requirements.txt        # Container Python dependencies
├── pyproject.toml          # PyPI package config for the MCP client
├── tests/                  # unit, live and integration suites
├── vpn/                    # mount point for VPN configs
└── CLAUDE.md               # maintainer and operational notes

Development

python -m pytest -q                     # 1506 passed, 6 skipped
python -m pytest -m live -q             # 15 passed, 3 skipped; needs a backend on :5000
python scripts/mutation_check.py --spec tests/mutations.json   # 103/103 guards verified red
python tests/integration/probe_tools.py # calls all 135 tools; needs a backend

pytest -m live exercises real execution rather than a mocked client: background job state across separate MCP processes, /etc/hosts round-trips, nmap and fingerprinting against a lab target, verbatim command output. Each case opens its own MCP stdio session, so any state surviving between calls is provably server-side. It skips itself when no backend answers, which keeps CI green without Docker — so a green pytest -q alone proves nothing about tool execution. Override KALI_API_URL, ZKM_LAB_HOST and ZKM_LAB_PORT to point it elsewhere; from inside the container the host lab is host.docker.internal.

probe_tools.py calls every tool once and diffs the outcome against tests/integration/probe_baseline.json, so a run prints only what changed. It is deliberately manual: real scanners, real listeners, and the public internet for a handful of OSINT tools the baseline marks best-effort. A raw BROKEN count is not a pass criterion — a tool truthfully reporting that no VPN is configured looks the same as one that regressed, and only the baseline tells them apart. Re-record it when a tool's expected outcome legitimately changes.

Live qualification fixtures

python tests/integration/run_smoke.py --image zebbern-kali-mcp:goal-full --network-mode bridge --expect-variant full
python tests/integration/run_smoke.py --image zebbern-kali-mcp:goal-full --network-mode host  --expect-variant full
python tests/integration/run_ad_lab.py --image zebbern-kali-mcp:goal-lean

Add --check-trim to run_smoke.py to also assert the live trim profile against the running image: it must expose 125 tools and omit exactly the nine callback_* tools plus parse_tool_output, with nothing else added or lost. It is opt-in because it costs one extra MCP session, and it is independent of the image variant, since trim is static and ignores capability discovery.

The AD fixture is local, disposable and publishes no host ports. It proves local DNS, authenticated LDAP discovery and the public MCP/API enumeration path, and nothing else about Active Directory operations.

Updating pinned build inputs

  1. Resolve an authoritative upstream version or commit.

  2. Update one Docker argument and its checksum when present.

  3. Run the Docker contract tests and docker build --check ..

  4. Build full and lean with CADO-NFS enabled.

  5. Run the image, bridge, AD, native-Linux host-network and Windows Docker Desktop host-network smoke fixtures.

  6. Compare tool names, image IDs and sizes, and common layers before accepting.

Releasing

Two release tracks: the wheel ships mcp_server.py and mcp_tools/*, while everything under zebbern-kali/ ships in the Docker image, so a backend change is not delivered by a PyPI release. The version lives in pyproject.toml and zebbern-kali/core/config.py and must be bumped in both (a test fails if only one moves). CLAUDE.md has the full procedure.


Contributing

Pull requests welcome — include a clear summary of the change and any relevant test notes.


Built on the Model Context Protocol · Created by Zebbern

Available Tools

131 tools
ad_asreproastA

Perform AS-REP Roasting to get hashes for accounts with pre-auth disabled.

Args: domain: AD domain dc_ip: Domain Controller IP username: Specific username to test userlist: Path to username list file on Kali

Inputs: domain (required), dc_ip, userlist, username. Returns: result

ad_bloodhound_collectB

Run BloodHound data collection against an Active Directory domain.

Args: domain: Target AD domain (e.g., corp.local) username: Domain username password: Domain password dc_ip: Domain Controller IP (auto-detected if empty) collection_method: Collection method (all, default, DCOnly, etc.) nameserver: Custom DNS nameserver

Inputs: domain (required), password (required), username (required), dc_ip, +2 more. Returns: result

ad_kerberoastC

Perform Kerberoasting attack to extract service ticket hashes.

Args: domain: AD domain username: Domain username password: Domain password dc_ip: Domain Controller IP target_user: Specific user to target (all SPNs if empty)

Inputs: domain (required), password (required), username (required), dc_ip, +1 more. Returns: result

ad_ldap_enumA

Enumerate Active Directory via LDAP.

Args: domain: AD domain username: Domain username password: Domain password dc_ip: Domain Controller IP or hostname query: Query type (users, groups, computers, spns, admins, all) use_starttls: Upgrade the LDAP connection to TLS tls_verify: Verify the StartTLS server certificate

Inputs: domain (required), password (required), username (required), dc_ip, +3 more. Returns: result

ad_password_sprayB

Perform password spraying attack against AD accounts.

Args: domain: AD domain password: Password to spray dc_ip: Domain Controller IP userlist: Path to username list file on Kali delay: Delay between attempts in seconds (default: 0)

Inputs: domain (required), password (required), dc_ip, delay, +1 more. Returns: result

ad_psexecC

Execute commands on a remote Windows host via PsExec (impacket).

Args: target: Target IP or hostname domain: AD domain username: Username password: Password hashes: NTLM hash for pass-the-hash command: Command to execute (default: cmd.exe)

Inputs: domain (required), target (required), username (required), hashes, +2 more. Returns: result

ad_secretsdumpB

Dump secrets (NTLM hashes, Kerberos keys) from a domain controller using secretsdump.py.

Args: domain: AD domain username: Domain admin username password: Domain admin password target: Specific target (default: DC) dc_ip: Domain Controller IP hashes: NTLM hash for pass-the-hash (LMHASH:NTHASH)

Inputs: domain (required), password (required), username (required), dc_ip, +2 more. Returns: result

ad_smb_enumC

Enumerate SMB shares and information on a target.

Args: target: Target IP or hostname username: Username for authenticated enum password: Password domain: AD domain

Inputs: target (required), domain, password, username. Returns: result

ad_tools_statusA

Check which AD tools are available on the Kali server.

Inputs: no named parameters. Returns: result

ad_wmiexecB

Execute commands on a remote Windows host via WMI (impacket).

Args: target: Target IP or hostname domain: AD domain username: Username password: Password hashes: NTLM hash for pass-the-hash command: Command to execute

Inputs: domain (required), target (required), username (required), hashes, +2 more. Returns: result

api_auth_bypass_testC

Test for authentication bypass vulnerabilities.

Args: url: Target API endpoint method: HTTP method headers: Custom headers as key:value pairs

Inputs: url (required), method, headers. Returns: result

api_ffuf_fuzzB

Fuzz using ffuf for content discovery and parameter brute-forcing.

Args: url: Target URL with FUZZ keyword (e.g., http://target/FUZZ) wordlist: Wordlist path on Kali server method: HTTP method mc: Match HTTP status codes (comma-separated) headers: Custom headers (key:value, comma-separated) data_str: POST data with FUZZ keyword

Inputs: url (required), mc, method, headers, +2 more. Returns: result

api_fuzz_endpointB

Fuzz an API endpoint with various payloads.

Args: url: Target API endpoint URL method: HTTP method (GET, POST, PUT, DELETE) parameters: Comma-separated parameter names to fuzz wordlist: Custom wordlist path (default: built-in) headers: Custom headers as key:value pairs, comma-separated

Inputs: url (required), method, headers, wordlist, +1 more. Returns: result

api_graphql_fuzzB

Fuzz a GraphQL endpoint for vulnerabilities.

Args: url: GraphQL endpoint URL query: Specific query to fuzz (auto-generated from schema if empty) depth: Fuzzing depth (default: 3)

Inputs: url (required), depth, query. Returns: result

api_graphql_introspectB

Introspect a GraphQL endpoint to discover schema, types, and queries.

Args: url: GraphQL endpoint URL

Inputs: url (required). Returns: result

api_jwt_analyzeA

Analyze a JWT token for weaknesses.

Args: token: The JWT token string to analyze

Returns: Decoded header/payload, algorithm analysis, and potential vulnerabilities

Inputs: token (required). Returns: result

api_jwt_crackB

Attempt to crack a JWT token's signing secret.

Args: token: The JWT token to crack wordlist: Path to wordlist on the Kali server

Inputs: token (required), wordlist. Returns: result

api_kiterunner_scanB

Scan API endpoints using Kiterunner for route discovery.

Args: url: Target base URL wordlist: Custom wordlist or kiterunner routes file

Inputs: url (required), wordlist. Returns: result

api_newman_runC

Run a Postman/Newman collection against API endpoints.

Args: collection: Path to Postman collection JSON file environment: Path to Postman environment JSON file

Inputs: collection (required), environment. Returns: result

api_nuclei_scanC

Run Nuclei templates against API endpoints.

Args: url: Target URL tags: Nuclei template tags (default: api) severity: Filter by severity (critical, high, medium, low)

Inputs: url (required), tags, severity. Returns: result

api_rate_limit_testA

Test API rate limiting controls.

Args: url: Target API endpoint requests_count: Number of requests to send (default: 100) method: HTTP method (default: GET)

Inputs: url (required), method, requests_count. Returns: result

callback_checkA

Check if any callbacks matching an identifier have been received.

Searches HTTP paths and DNS query names for the given substring. Useful after injecting a payload to see if the target called back.

Args: identifier: Substring to match in the request path or DNS query name. Empty string matches all callbacks. since_minutes: Only check callbacks from the last N minutes (default 60).

Returns: Dict with found status, match count, and matching entries.

Example: callback_check(identifier='ssrf-test1') callback_check(identifier='xxe', since_minutes=5)

Inputs: identifier, since_minutes. Returns: result

callback_clearA

Clear all captured callbacks from the catcher's memory.

Removes all stored HTTP and DNS callback entries. The listeners continue running — only the stored data is cleared.

Returns: Confirmation with the count of cleared entries.

Example: callback_clear()

Inputs: no named parameters. Returns: result

callback_generateA

Generate callback payload URLs and commands for injection testing.

Creates ready-to-use payloads with unique identifiers that will be captured by the callback catcher when triggered.

Args: listener_ip: IP address of the callback listener (your Kali IP). http_port: HTTP listener port (default 8888). dns_port: DNS listener port (default 5353). payload_type: Type of payload to generate. Options: 'url' — simple callback URL 'curl' — curl and wget one-liners 'xxe' — XML External Entity payloads 'ssrf' — SSRF test URLs 'dns' — DNS lookup commands (nslookup, dig, host) 'all' — all of the above

Returns: Generated payloads with unique identifiers for tracking.

Example: callback_generate(listener_ip='10.10.14.5') callback_generate(listener_ip='10.10.14.5', payload_type='xxe')

Inputs: listener_ip (required), dns_port, http_port, payload_type. Returns: result

callback_latestA

Get the most recent callback captured by the local catcher.

Returns the single latest callback entry, whether HTTP or DNS. Useful for quick checks after sending a payload.

Returns: The latest callback dict with full request details, or a message if no callbacks have been captured.

Example: callback_latest()

Inputs: no named parameters. Returns: result

callback_listA

List captured callbacks from the local catcher.

Returns the most recent callbacks, newest first. Filter by type to see only HTTP requests or DNS queries.

Args: limit: Maximum number of entries to return (default 50). callback_type: Filter by 'http', 'dns', or 'all' (default 'all').

Returns: List of callback entries with timestamps, source IPs, request details, and bodies.

Example: callback_list() callback_list(limit=10, callback_type='dns')

Inputs: limit, callback_type. Returns: result

callback_startA

Start the local callback catcher (HTTP + DNS listeners).

Runs an HTTP server and a DNS server on the Kali container to capture out-of-band callbacks from targets. Use this instead of webhook.site when the target network has no internet access.

Args: http_port: TCP port for the HTTP listener (default 8888). dns_port: UDP port for the DNS listener (default 5353). bind_ip: IP address to bind to (default '0.0.0.0'). Use the tun0 IP to restrict to VPN traffic.

Returns: Listener status including ports, bind IP, and detected tun0 IP.

Example: callback_start() callback_start(http_port=9090, bind_ip='10.10.14.5')

Inputs: bind_ip, dns_port, http_port. Returns: result

callback_statusA

Get the current status of the callback catcher.

Returns whether the listeners are running, which ports they are on, and counts of captured HTTP and DNS callbacks.

Returns: Status dict with running state, ports, and callback counts.

Example: callback_status()

Inputs: no named parameters. Returns: result

callback_stopA

Stop the local callback catcher and clean up all listeners.

Shuts down both the HTTP and DNS listener threads and closes their sockets. Captured callbacks remain in memory until explicitly cleared or a new session starts.

Returns: Confirmation with the number of callbacks still in memory.

Example: callback_stop()

Inputs: no named parameters. Returns: result

callback_waitA

Wait for a new callback to arrive at the local catcher.

Polls the callback catcher until a new entry appears or the timeout is reached. Checks every 2 seconds.

Args: timeout_seconds: Maximum time to wait in seconds (default 60). callback_type: Wait for a specific type: 'http', 'dns', or 'all' (default 'all').

Returns: The new callback entry if one arrived, or a timeout message.

Example: callback_wait() callback_wait(timeout_seconds=120, callback_type='http')

Inputs: callback_type, timeout_seconds. Returns: result

ctf_connectA

Connect to a CTF platform and authenticate.

Supports CTFd (most common) and rCTF platforms. Must be called before using other ctf_* tools.

Args: url: CTF platform base URL (e.g. 'https://ctf.example.com') token: API token from the platform settings page. For CTFd: Settings → Access Tokens → Generate. For rCTF: Profile → API Token. platform_type: Platform type — 'ctfd' (default) or 'rctf' verify_ssl: Verify SSL certificates (default True). Set False for self-signed certs commonly used in local CTF deployments.

Example: ctf_connect(url='https://ctf.example.com', token='ctfd_abc123')

Inputs: url (required), token, verify_ssl, platform_type. Returns: result

ctf_download_fileA

Download challenge files from the CTF platform.

Either provide challenge_id (downloads the first file) or a direct file_url from the challenge details.

Args: challenge_id: Download files for this challenge (uses first file) file_url: Direct file URL (takes priority over challenge_id) output_dir: Directory to save files to (default: /app/tmp/ctf_files)

Inputs: file_url, output_dir, challenge_id. Returns: result

ctf_get_challengeA

Get full details for a specific CTF challenge.

Returns the challenge description, files, hints, connection info, and solve status.

Args: challenge_id: The numeric challenge ID (from ctf_list_challenges)

Inputs: challenge_id (required). Returns: result

ctf_list_challengesA

List all available CTF challenges.

Returns challenge names, categories, point values, and solve counts. Requires a prior ctf_connect() call.

Args: category: Optional category filter (e.g. 'web', 'crypto', 'forensics')

Inputs: category. Returns: result

ctf_scoreboardA

Fetch the current CTF scoreboard.

Args: top: Number of top entries to return (default 20)

Inputs: top. Returns: result

ctf_statusA

Check current CTF platform connection status.

Inputs: no named parameters. Returns: result

ctf_submit_flagA

Submit a flag for a CTF challenge.

Returns whether the flag was correct, already solved, or incorrect.

Args: challenge_id: The challenge ID to submit for flag: The flag string (e.g. 'flag{example_flag_here}')

Example: ctf_submit_flag(challenge_id=42, flag='flag{s0m3_fl4g}')

Inputs: flag (required), challenge_id (required). Returns: result

cve_package_auditA

Check a specific software package for known vulnerabilities using OSV.dev.

Queries the Open Source Vulnerabilities database to find all known CVEs and security advisories affecting a package. Supports version-specific queries across all major ecosystems.

Args: package: Package name (e.g., 'lodash', 'django', 'openssl') version: Specific version to check (e.g., '4.17.20'). If omitted, returns all known vulnerabilities for any version. ecosystem: Package ecosystem — 'npm', 'PyPI', 'Maven', 'Go', 'crates.io', 'NuGet', 'Packagist', 'RubyGems', 'Debian', 'Alpine', etc. If omitted, OSV auto-detects.

Example usage: cve_package_audit(package='lodash', version='4.17.20', ecosystem='npm') cve_package_audit(package='django', ecosystem='PyPI') cve_package_audit(package='openssl')

Inputs: package (required), version, ecosystem. Returns: result

exec_streamA

Execute a command with real-time streaming output via SSE (text/event-stream). Posts to api/command with streaming=True. Useful for long-running commands like nmap, nuclei, fuzzing.

Args: command: The command to execute timeout: Timeout in seconds (default: 3600 = 1 hour)

Returns: Streaming output collected in real-time with all events

Inputs: command (required), timeout. Returns: result

exploit_copyA

Copy an exploit to a working directory for modification.

Args: edb_id: Exploit-DB ID number destination: Directory to copy to (default: /tmp)

Returns: Path to copied exploit file

Inputs: edb_id (required), destination. Returns: result

exploit_detailsA

Get full details and source code for an exploit.

Args: edb_id: Exploit-DB ID number

Returns: Exploit metadata and source code

Inputs: edb_id (required). Returns: result

exploit_suggest_for_serviceB

Get exploit suggestions for a specific service.

Args: service: Service name (e.g., 'apache', 'openssh', 'vsftpd') version: Version string (optional, e.g., '2.3.4', '8.4p1')

Returns: List of suggested exploits

Inputs: service (required), version. Returns: result

exploit_suggest_from_nmapC

Analyze nmap scan output and suggest exploits for discovered services.

Args: nmap_output: Raw nmap scan output text

Returns: Exploit suggestions grouped by service/port

Inputs: nmap_output (required). Returns: result

fingerprint_headersA

Analyze HTTP response headers for security posture.

Args: url: Target URL to check headers

Returns: Header analysis with security ratings

Inputs: url (required). Returns: result

fingerprint_urlA

Fingerprint a URL to detect technologies, frameworks, and CMS.

Args: url: Target URL to fingerprint

Returns: Detected technologies, headers, and server information

Inputs: url (required). Returns: result

fingerprint_wafB

Detect Web Application Firewall (WAF) on a target URL.

Args: url: Target URL to test for WAF

Returns: WAF detection results

Inputs: url (required). Returns: result

healthA

Check the health status of the Kali API server.

Returns: Server health information

Inputs: no named parameters. Returns: result

hosts_addA

Add hostname(s) to /etc/hosts on the Kali container at runtime.

No container restart needed. Entries are tracked in a managed section so they can be listed, removed, or cleared independently of system entries.

Args: ip: Target IP address (e.g., '10.129.17.154') hostnames: Space or comma separated hostnames (e.g., 'target.htb www.target.htb dc01.corp.local')

Returns: Added entries and current state

Example: hosts_add(ip='10.129.17.154', hostnames='silentium.htb') hosts_add(ip='10.10.11.5', hostnames='dc01.corp.local corp.local')

Inputs: ip (required), hostnames (required). Returns: result

hosts_clearA

Remove ALL managed /etc/hosts entries.

Clears only the Kali-MCP managed section. System entries (localhost, etc.) are never touched. Useful when switching to a completely new target.

Returns: Count of removed entries

Example: hosts_clear()

Inputs: no named parameters. Returns: result

hosts_listA

List all managed /etc/hosts entries added via Kali-MCP.

Returns: List of managed entries with IP and hostnames

Example: hosts_list()

Inputs: no named parameters. Returns: result

hosts_removeA

Remove a hostname from managed /etc/hosts entries.

Only removes from the Kali-MCP managed section — system entries (localhost, etc.) are never touched.

Args: hostname: The hostname to remove (e.g., 'target.htb')

Returns: Removal result and remaining entries

Example: hosts_remove(hostname='old-target.htb')

Inputs: hostname (required). Returns: result

job_cancelA

Cancel a running background job and its child process group.

Inputs: job_id (required). Returns: result

job_outputC

Poll recent bounded stdout and stderr from a background job.

Inputs: job_id (required), lines, timeout. Returns: result

job_statusA

Return state and exit metadata for a background command job.

Inputs: job_id (required). Returns: result

kali_downloadA

Download file content from the Kali server as base64.

Args: remote_path: Path to file on the Kali server verify_checksum: Verify SHA256 checksum if provided by the server

Returns: File content encoded as base64

Inputs: remote_path (required), verify_checksum. Returns: result

kali_uploadB

Upload content to the Kali server filesystem.

Args: content: Base64-encoded file content remote_path: Destination path on the Kali server encoding: Content encoding (utf-8, binary) verify_checksum: Compute and send SHA256 checksum for integrity verification

Inputs: content (required), remote_path (required), encoding, verify_checksum. Returns: result

msf_session_createA

Create a new persistent Metasploit (msfconsole) session.

Returns: Session ID to use with other msf_session_* tools

Inputs: no named parameters. Returns: result

msf_session_destroyB

Destroy a specific Metasploit session.

Args: session_id: The session ID to destroy

Inputs: session_id (required). Returns: result

msf_session_destroy_allA

Destroy all active Metasploit sessions.

Inputs: no named parameters. Returns: result

msf_session_executeA

Execute a command in an existing Metasploit session.

Args: session_id: The session ID from msf_session_create command: The Metasploit command to execute (e.g., "use exploit/...", "set RHOSTS ...", "run") timeout: Command timeout in seconds (default: 300) read_delay: Seconds to wait before reading output for slow-responding commands (default: 2)

Returns: Command output and status

Inputs: command (required), session_id (required), timeout, read_delay. Returns: result

msf_session_listA

List all active Metasploit sessions.

Inputs: no named parameters. Returns: result

parse_tool_outputA

Parse raw security tool output into structured JSON.

Takes the raw text output from tools like nmap, nuclei, or gobuster and returns a structured JSON representation for easier analysis. Parsing happens locally — no backend call is made.

Args: output: Raw text output from the tool tool_name: Tool identifier — 'nmap', 'nuclei', 'gobuster', or 'auto' output_format: Format hint — 'xml', 'jsonl', 'text', or 'auto' (default: auto)

Inputs: output (required), tool_name (required), output_format. Returns: result

payload_generateA

Generate a payload using msfvenom.

Args: lhost: Your IP address for callback lport: Port for callback (default: 4444) template: Use a predefined template (e.g., 'windows_meterpreter_reverse_tcp') payload: Metasploit payload string (ignored if template specified) format_type: Output format - exe, elf, raw, ps1, etc. encoder: Encoder to use (e.g., 'x86/shikata_ga_nai') iterations: Number of encoding iterations bad_chars: Characters to avoid in payload (e.g., '\x00\x0a\x0d') nops: Number of NOP sled bytes to prepend (default: 0) output_name: Custom output filename (auto-generated if empty)

Returns: Generated payload info including base64 content for small payloads

Inputs: lhost (required), nops, lport, encoder, +6 more. Returns: result

payload_host_startB

Start HTTP server to host generated payloads for download.

Args: port: Port for the hosting server (default: 8888)

Returns: Server URL for downloading payloads

Inputs: port. Returns: result

payload_host_stopA

Stop the payload hosting server.

Inputs: no named parameters. Returns: result

payload_listB

List all generated payloads.

Inputs: no named parameters. Returns: result

payload_one_linerB

Generate reverse shell one-liner commands.

Args: lhost: Your IP address for callback lport: Port for callback (default: 4444) shell_type: Type of shell - bash, python, nc, php, powershell, perl, ruby, or 'all'

Returns: One-liner command(s) ready to copy-paste

Inputs: lhost (required), lport, shell_type. Returns: result

payload_templatesA

List available payload templates and encoders for msfvenom.

Returns: List of template names and available encoders

Inputs: no named parameters. Returns: result

pivot_add_pivotB

Add a pivot point for tracking network access chains.

Args: name: Pivot name/identifier pivot_host: Host used as pivot method: Pivot method (ssh, chisel, ligolo, socat) subnet: Reachable subnet via this pivot (e.g., 10.10.10.0/24) notes: Additional notes

Inputs: name (required), pivot_host (required), notes, method, +1 more. Returns: result

pivot_chisel_clientA

Start a Chisel client to connect to a Chisel server.

Args: server_url: Chisel server URL (e.g., http://attacker:8080) tunnels: Tunnel specifications (e.g., 'R:8888:10.10.10.5:80') fingerprint: Server fingerprint for verification

Inputs: tunnels (required), server_url (required), fingerprint. Returns: result

pivot_chisel_serverB

Start a Chisel server for tunneling.

Args: port: Listening port (default: 8080) reverse: Allow reverse tunnels (default: True)

Inputs: port, reverse. Returns: result

pivot_generate_proxychainsA

Generate a proxychains configuration for pivoting through SOCKS proxy.

Args: socks_port: SOCKS proxy port (default: 1080) proxy_type: Proxy type (socks4, socks5)

Returns: Proxychains configuration content

Inputs: proxy_type, socks_port. Returns: result

pivot_ligolo_startC

Start a Ligolo-ng proxy server for pivoting.

Args: interface: TUN interface name (default: ligolo) port: Listening port (default: 11601)

Inputs: port, interface. Returns: result

pivot_list_pivotsB

List all configured pivot points/routes.

Inputs: no named parameters. Returns: result

pivot_list_tunnelsA

List all active tunnels and port forwards.

Inputs: no named parameters. Returns: result

pivot_socat_forwardB

Create a socat port forward.

Args: listen_port: Local port to listen on target_host: Target host to forward to target_port: Target port to forward to protocol: Protocol (tcp, udp)

Inputs: listen_port (required), target_host (required), target_port (required), protocol. Returns: result

pivot_ssh_dynamicB

Create an SSH dynamic SOCKS proxy.

Args: ssh_host: SSH server to use as SOCKS proxy socks_port: Local SOCKS port (default: 1080) username: SSH username password: SSH password key_file: SSH key file path

Inputs: ssh_host (required), key_file, password, username, +1 more. Returns: result

pivot_ssh_localA

Create an SSH local port forward (access remote service locally).

Args: ssh_host: SSH server to tunnel through local_port: Local port to listen on remote_host: Remote host to forward to (from SSH server's perspective) remote_port: Remote port to forward to username: SSH username password: SSH password key_file: SSH key file path

Inputs: ssh_host (required), local_port (required), remote_host (required), remote_port (required), +3 more. Returns: result

pivot_ssh_remoteA

Create an SSH remote port forward (expose local service on remote host).

Args: ssh_host: SSH server to tunnel through remote_port: Port to open on the SSH server local_host: Local host to forward to local_port: Local port to forward to username: SSH username password: SSH password key_file: SSH key file path

Inputs: ssh_host (required), local_host (required), local_port (required), remote_port (required), +3 more. Returns: result

pivot_stop_all_tunnelsA

Stop all active tunnels and port forwards.

Inputs: no named parameters. Returns: result

pivot_stop_tunnelB

Stop a specific tunnel.

Args: tunnel_id: The tunnel ID to stop

Inputs: tunnel_id (required). Returns: result

read_outputA

Read or poll bounded output from a background command job.

Typical workflow:

  1. send_input(session_id, "whoami\n")

  2. read_output(session_id, timeout=5) -> returns the command's output

The backend will wait up to timeout seconds for new output before returning whatever is available (which may be empty if the process has not produced anything yet).

Args: session_id: The session identifier to read from. timeout: Maximum seconds the backend should wait for new output before returning (default: 5). Use a higher value for slow commands (e.g. nmap, compilation). lines: Maximum number of output lines to return (default: 100). Older lines are trimmed first when the buffer exceeds this.

Returns: dict with at minimum: - success (bool): whether the read succeeded - output (str): the collected output text - session_id (str): echo of the session targeted - lines_returned (int): number of lines in output - error (str, optional): present only on failure

Inputs: session_id (required), lines, timeout. Returns: result

reverse_shell_commandA

Execute a command in an active reverse shell session.

Args: session_id: The session ID (e.g., 'shell_4444') command: Command to execute on the target timeout: Command timeout in seconds

Returns: Command output from the target system

Inputs: command (required), session_id (required), timeout. Returns: result

reverse_shell_download_contentA

Download file content from target via reverse shell and return as base64.

Args: session_id: The reverse shell session ID remote_file: Path to the file on the target method: Download method (base64, cat)

Inputs: session_id (required), remote_file (required), method. Returns: result

reverse_shell_generate_payloadA

Generate reverse shell payloads for manual execution on targets.

Args: local_ip: Your local IP address that the target should connect back to local_port: Local port to connect back to (default: 4444) payload_type: Type of payload (bash, python, nc, php, powershell, perl) encoding: Encoding format (plain, base64, url, hex)

Returns: Generated payload in various formats ready for manual execution

Inputs: local_ip (required), encoding, local_port, payload_type. Returns: result

reverse_shell_listener_startA

Start a reverse shell listener on the specified port.

Args: port: Port to listen on (default: 4444) session_id: Optional session identifier (auto-generated as shell_{port} if empty) listener_type: Type of listener - 'netcat' or 'pwncat' (default: netcat) auto_upgrade: Automatically attempt TTY upgrade on connection (default: False)

Returns: Session ID and listener status

Inputs: port, session_id, auto_upgrade, listener_type. Returns: result

reverse_shell_send_payloadA

Send a payload command to trigger a reverse shell connection in a non-blocking way.

Executes the payload in a background thread to avoid blocking the server. Waits then returns session status to verify the connection was established.

Args: session_id: The session ID of the reverse shell listener payload_command: The payload command to execute (e.g., curl with reverse shell) timeout: Timeout for the payload execution in seconds (default: 10) wait_seconds: Seconds to wait before checking session status (default: 5)

Returns: Payload execution status and session info

Inputs: session_id (required), payload_command (required), timeout, wait_seconds. Returns: result

reverse_shell_statusA

Get the status of reverse shell sessions.

Args: session_id: Optional specific session ID to check (if empty, shows all sessions)

Returns: Status information for reverse shell sessions

Inputs: session_id. Returns: result

reverse_shell_stopB

Stop a reverse shell session.

Args: session_id: The session ID to stop

Returns: Stop operation result

Inputs: session_id (required). Returns: result

reverse_shell_upload_contentB

Upload content directly to the target via reverse shell.

Args: session_id: The reverse shell session ID content: Base64 encoded content to upload remote_file: Path where to save the file on the target method: Upload method (base64) encoding: Content encoding (utf-8, binary)

Inputs: content (required), session_id (required), remote_file (required), method, +1 more. Returns: result

send_inputA

Send text input to a running background command job.

Use this together with read_output() to have a full interactive conversation with a long-running process:

  1. Start a job with zebbern_exec(..., background=True)

  2. send_input(session_id, "some command\n")

  3. read_output(session_id) to collect the response

Args: session_id: The job identifier returned by zebbern_exec. input_text: The text to send to the session's stdin. Include a trailing newline (\n) if the target process expects one. session_type: Compatibility hint retained for existing clients.

Returns: dict with at minimum: - success (bool): whether the input was accepted - session_id (str): echo of the session targeted - error (str, optional): present only on failure

Inputs: input_text (required), session_id (required), session_type. Returns: result

ssh_estimate_transferB

Estimate file transfer time over SSH.

Args: file_size_bytes: File size in bytes operation: Transfer operation type (upload, download)

Returns: Estimated transfer time and recommendations

Inputs: file_size_bytes (required), operation. Returns: result

ssh_session_commandC

Execute a command in an active SSH session.

Args: session_id: The SSH session ID command: Command to execute on the remote host timeout: Command timeout in seconds (default: 60)

Returns: Command output from remote host

Inputs: command (required), session_id (required), timeout. Returns: result

ssh_session_download_contentB

Download file content from a remote host via SSH as base64.

Args: session_id: The SSH session ID remote_path: Path to the file on the remote host

Inputs: session_id (required), remote_path (required). Returns: result

ssh_sessionsB

List all active SSH sessions.

Inputs: no named parameters. Returns: result

ssh_session_startA

Start a persistent SSH session to a target host.

Args: target: Target hostname or IP username: SSH username (default: root) password: SSH password (if using password auth) key_file: Path to SSH private key file (if using key auth) port: SSH port (default: 22) session_id: Optional session identifier (auto-generated if empty)

Returns: Session ID and connection status

Inputs: target (required), port, key_file, password, +2 more. Returns: result

ssh_session_statusA

Check the status of an SSH session.

Args: session_id: The SSH session ID to check

Inputs: session_id (required). Returns: result

ssh_session_stopB

Stop/disconnect an SSH session.

Args: session_id: The SSH session ID to stop

Inputs: session_id (required). Returns: result

ssh_session_upload_contentA

Upload content directly to a remote host via SSH (no local temp files).

Args: session_id: The SSH session ID content: Base64-encoded content to upload remote_path: Destination path on the remote host encoding: Content encoding (utf-8, binary)

Inputs: content (required), session_id (required), remote_path (required), encoding. Returns: result

system_network_infoA

Get comprehensive network information for the Kali Linux system.

Returns: Network information including interfaces, IP addresses, routing table, etc.

Inputs: no named parameters. Returns: result

target_download_fileB

Download file content from a target via an active session.

Args: session_id: Active session ID remote_path: Path to file on the target method: Transfer method (ssh, reverse_shell) verify_checksum: Verify SHA256 checksum if provided by the server

Returns: File content encoded as base64

Inputs: session_id (required), remote_path (required), method, verify_checksum. Returns: result

target_upload_fileA

Upload content to a target via an active session (SSH or reverse shell).

Args: session_id: Active session ID content: Base64-encoded file content remote_path: Destination path on the target method: Transfer method (ssh, reverse_shell) encoding: Content encoding (utf-8, binary) verify_checksum: Compute and send SHA256 checksum for integrity verification

Inputs: content (required), session_id (required), remote_path (required), method, +2 more. Returns: result

tools_amassB

Perform advanced subdomain enumeration using OWASP Amass.

Amass performs network mapping of attack surfaces and external asset discovery using open-source information gathering and active reconnaissance techniques.

Args: domain: Target domain for subdomain enumeration (e.g., 'example.com') mode: Enumeration mode — 'passive' for OSINT-only or 'active' for DNS resolution and zone transfers (default: 'passive') additional_args: Extra amass arguments

Example usage: tools_amass(domain='example.com') tools_amass(domain='target.org', mode='active')

Inputs: domain (required), mode, additional_args. Returns: result

tools_arjunB

Discover hidden HTTP parameters using Arjun.

Args: url: Target URL method: HTTP method (GET, POST, JSON) headers: Custom headers as key:value pairs, comma-separated wordlist: Custom wordlist path on Kali (default: Arjun built-in) delay: Delay between requests in seconds (default: 0) threads: Number of concurrent threads (default: 2) include: Only test these parameters (comma-separated) exclude: Skip these parameters (comma-separated)

Inputs: url (required), delay, method, exclude, +4 more. Returns: result

tools_assetfinderC

Execute Assetfinder for asset discovery.

Args: domain: Target domain additional_args: Extra assetfinder arguments

Inputs: domain (required), additional_args. Returns: result

tools_byp4xxC

Execute byp4xx for 403 bypass testing.

Args: url: Target URL returning 403 method: HTTP method additional_args: Extra arguments

Inputs: url (required), method, additional_args. Returns: result

tools_crtshA

Query crt.sh certificate transparency logs for passive subdomain enumeration.

Searches Certificate Transparency (CT) logs via crt.sh to discover subdomains that have had TLS certificates issued for them. This is a passive reconnaissance technique that does not send any traffic to the target domain itself.

Args: domain: Target domain to search (e.g., 'example.com') include_expired: Include expired certificates in results (default: False). When True, results may contain historical subdomains that are no longer active but once held valid certificates.

Example usage: tools_crtsh(domain='example.com') tools_crtsh(domain='target.org', include_expired=True)

Inputs: domain (required), include_expired. Returns: result

tools_enum4linuxC

Execute Enum4linux Windows/Samba enumeration tool.

Args: target: Target IP additional_args: Enum4linux flags (default: -a for all)

Inputs: target (required), additional_args. Returns: result

tools_fierceC

Execute Fierce for DNS reconnaissance.

Args: domain: Target domain additional_args: Extra fierce arguments

Inputs: domain (required), additional_args. Returns: result

tools_gobusterC

Execute Gobuster to find directories, DNS subdomains, or virtual hosts.

Args: url: Target URL mode: Gobuster mode (dir, dns, vhost, fuzz) wordlist: Wordlist path on Kali additional_args: Extra gobuster arguments threads: Number of concurrent threads (default: 10) extensions: File extensions to search for, comma-separated (e.g., 'php,html,txt') status_codes: Positive status codes to match, comma-separated (e.g., '200,301,302')

Inputs: url (required), mode, threads, wordlist, +3 more. Returns: result

tools_gowitnessA

Capture a screenshot of a web page using gowitness.

Uses a headless Chrome browser to render and screenshot web pages. Useful for visual reconnaissance, documenting findings, and identifying web technologies from page appearance.

Args: url: Target URL to screenshot (e.g., 'https://example.com') threads: Number of concurrent screenshot threads (default: 4) resolution: Browser viewport resolution as WIDTHxHEIGHT (default: '1280x720') additional_args: Extra gowitness arguments

Example usage: tools_gowitness(url='https://example.com') tools_gowitness(url='https://target.org', threads=8, resolution='1920x1080')

Inputs: url (required), threads, resolution, additional_args. Returns: result

tools_httpxB

Execute httpx for HTTP probing.

Args: target: Target URL, domain, or IP additional_args: Extra httpx arguments

Inputs: target (required), additional_args. Returns: result

tools_hydraB

Execute Hydra password cracking tool.

Args: target: Target IP or hostname service: Service to attack (ssh, ftp, http-get, etc.) username: Single username username_file: Path to username list on Kali password: Single password password_file: Path to password list on Kali additional_args: Extra hydra arguments tasks: Number of parallel connection threads (default: 16) wait: Timeout in seconds for each connection attempt (default: 32) port: Target port override (default: 0 means use service default)

Inputs: target (required), service (required), port, wait, +6 more. Returns: result

tools_johnA

Execute John the Ripper password cracker.

Args: hash_file: Path to hash file on Kali wordlist: Wordlist path format_type: Hash format (e.g., 'NT', 'md5crypt', 'raw-sha256') additional_args: Extra john arguments

Inputs: hash_file (required), wordlist, format_type, additional_args. Returns: result

tools_katanaA

Execute Katana for web crawling with JavaScript parsing support.

Katana is a next-generation crawling and spidering framework that supports headless browsing and JavaScript rendering, allowing it to discover endpoints hidden behind client-side routing and dynamically generated content.

Args: url: Target URL to crawl (e.g., 'https://example.com') depth: Maximum crawl depth (default: 3). Higher values discover more pages but take longer. js_crawl: Enable JavaScript crawling/rendering (default: True). When enabled, Katana uses a headless browser to execute JS and discover dynamically loaded endpoints. scope: Regex pattern to restrict crawl scope (e.g., '.*.example.com'). Empty string means no scope restriction. additional_args: Extra katana arguments (e.g., '-H "Authorization: Bearer token"')

Example usage: tools_katana(url='https://example.com') tools_katana(url='https://target.com', depth=5, js_crawl=True) tools_katana(url='https://app.example.com', scope='.*.example.com', additional_args='-H "Cookie: session=abc"')

Inputs: url (required), depth, scope, js_crawl, +1 more. Returns: result

tools_masscanA

Execute Masscan for fast port scanning across large IP ranges.

Masscan is an internet-scale port scanner capable of scanning the entire internet in under 6 minutes. It produces SYN packets asynchronously, making it significantly faster than traditional scanners for large target ranges.

Args: target: Target IP, CIDR range, or hostname (e.g., '10.0.0.0/24') ports: Port specification (default: '1-65535' for full range). Supports ranges ('1-1024'), lists ('80,443,8080'), or mixed ('22,80,443,8000-9000'). rate: Packets per second (default: 1000). Higher values scan faster but may overwhelm the network or trigger IDS. Common values: 1000 (safe), 10000 (moderate), 100000 (aggressive). additional_args: Extra masscan arguments (e.g., '--banners' to grab service banners)

Example usage: tools_masscan(target='192.168.1.0/24', ports='80,443,8080', rate=5000) tools_masscan(target='10.0.0.1', ports='1-1024') tools_masscan(target='172.16.0.0/16', ports='22,80,443', rate=10000, additional_args='--banners')

Inputs: target (required), rate, ports, additional_args. Returns: result

tools_niktoB

Execute Nikto web server scanner.

Args: target: Target URL or IP additional_args: Extra nikto arguments tuning: Nikto scan tuning options (e.g., '1' for interesting file, '2' for misconfiguration) output_format: Output format (e.g., 'htm', 'csv', 'xml', 'txt')

Inputs: target (required), tuning, output_format, additional_args. Returns: result

tools_nmapA

Execute an Nmap scan against a target.

Args: target: Target IP, hostname, or CIDR range scan_type: Nmap scan type flags (default: -sV) ports: Port specification (e.g., '80,443' or '1-1024') additional_args: Extra nmap arguments output_format: Output format — 'normal', 'xml', or 'grepable' (default: normal). When 'xml', adds -oX - for structured XML output.

Inputs: target (required), ports, scan_type, output_format, +1 more. Returns: result

tools_sqlmapB

Execute SQLmap SQL injection scanner.

Args: url: Target URL with parameter(s) data: POST data string additional_args: Extra sqlmap arguments technique: SQL injection techniques to test (e.g., 'BEUSTQ') level: Level of tests to perform, 1-5 (default: 1) risk: Risk of tests to perform, 1-3 (default: 1) dbs: Enumerate DBMS databases (default: False) tables: Enumerate DBMS database tables (default: False) dump: Dump DBMS database table entries (default: False)

Inputs: url (required), dbs, data, dump, +5 more. Returns: result

tools_ssh_auditB

Execute ssh-audit to analyze SSH server security configuration.

Args: target: Target hostname or IP port: SSH port (default: 22) timeout: Connection timeout in seconds scan_type: Scan type (ssh1, ssh2) policy_file: Path to policy file on Kali additional_args: Extra arguments

Inputs: target (required), port, timeout, scan_type, +2 more. Returns: result

tools_sslscanA

Execute sslscan to test SSL/TLS ciphers and certificate configuration.

Sslscan probes a target's SSL/TLS service to enumerate supported cipher suites, protocol versions, and certificate details. Useful for identifying weak ciphers, expired certificates, and misconfigured TLS deployments.

Args: target: Target hostname or IP to scan (e.g., 'example.com') port: TLS port to connect to (default: 443) additional_args: Extra sslscan arguments (e.g., '--no-colour' or '--show-certificate')

Example usage: tools_sslscan(target='example.com') tools_sslscan(target='10.0.0.5', port=8443) tools_sslscan(target='mail.example.com', port=993, additional_args='--show-certificate')

Inputs: target (required), port, additional_args. Returns: result

tools_subfinderB

Execute Subfinder for subdomain enumeration.

Args: target: Target domain additional_args: Extra subfinder arguments

Inputs: target (required), additional_args. Returns: result

tools_subzyC

Execute Subzy for subdomain takeover detection.

Args: target: Target domain or file with subdomains additional_args: Extra subzy arguments

Inputs: target (required), additional_args. Returns: result

tools_waybackurlsC

Execute waybackurls to fetch URLs from Wayback Machine.

Args: domain: Target domain additional_args: Extra arguments

Inputs: domain (required), additional_args. Returns: result

tools_wpscanA

Execute WPScan WordPress vulnerability scanner.

Args: url: Target WordPress URL additional_args: Extra wpscan arguments api_token: WPScan API token for vulnerability data lookups enumerate: Enumeration options (e.g., 'vp' for vulnerable plugins, 'u' for users) output_format: Output format — 'cli', 'json', or 'cli-no-colour' (default: cli)

Inputs: url (required), api_token, enumerate, output_format, +1 more. Returns: result

vpn_connectA

Connect to a VPN (WireGuard or OpenVPN) with a single call.

Places the config, brings the tunnel up, and returns the assigned IP address. Auto-detects VPN type from the config file contents unless explicitly specified.

Args: config_path: Absolute path to the VPN config file inside the container (e.g. /vpn/wg0.conf or /vpn/client.ovpn). Mount your config directory as a Docker volume to /vpn. vpn_type: 'wireguard', 'openvpn', or 'auto' (default). Auto reads [Interface]/[Peer] for WireGuard, or client/remote directives for OpenVPN. interface: WireGuard interface name (default wg0). Ignored for OpenVPN connections.

Returns: Connection result including success flag, assigned IP, interface name, and VPN type.

Example: vpn_connect(config_path='/vpn/wg0.conf') vpn_connect(config_path='/vpn/client.ovpn', vpn_type='openvpn')

Inputs: config_path (required), vpn_type, interface. Returns: result

vpn_disconnectA

Disconnect from a VPN tunnel.

Brings down the specified interface (WireGuard) or kills the OpenVPN daemon process.

Args: interface: Interface to disconnect (default wg0). vpn_type: 'wireguard', 'openvpn', or 'auto'. Auto checks whether the interface is a WireGuard tunnel.

Returns: Disconnection result with success flag.

Example: vpn_disconnect() vpn_disconnect(interface='wg1') vpn_disconnect(vpn_type='openvpn')

Inputs: vpn_type, interface. Returns: result

vpn_statusA

Get status of all active VPN connections.

Queries both WireGuard (via wg show) and OpenVPN (via PID file) and returns a list of active tunnels with their assigned IP addresses.

Returns: Dictionary with connections list, count, and active_count.

Example: vpn_status()

Inputs: no named parameters. Returns: result

zebbern_execA

Execute ANY command on the Kali server without restrictions. Full root access, no timeout limits (default 1 hour).

Args: command: The command to execute (can be any shell command, pipes, chains, etc.) timeout: Timeout in seconds (default: 3600 = 1 hour) cwd: Optional working directory for the command background: If True, return immediately with a trackable job_id

Returns: Command output with stdout, stderr, return_code, execution_time. When background=True, returns job_id, pid, and initial status instead.

Inputs: command (required), cwd, timeout, background. Returns: result

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    C2.9/5.0

    Scored across 131 tools

    Disambiguation2/5

    Multiple tools target the same job—subdomain discovery has subfinder/amass/assetfinder/crtsh/fierce, API fuzzing has api_fuzz_endpoint/api_ffuf_fuzz/api_nuclei_scan/api_kiterunner_scan, and payload generation has payload_generate/payload_one_liner/reverse_shell_generate_payload. Four different session/upload families add further overlap. Agents will frequently need to inspect descriptions closely to avoid selecting the wrong tool.

    Naming Consistency2/5

    Naming is clustered by feature area but not unified; the server mixes noun-first names like tools_nmap and payload_list with verb-first names like exec_stream and callback_generate. The prefix scheme helps within categories, but cross-category patterns and generic names like zebbern_exec and health make the overall naming inconsistent.

    Tool Count1/5

    131 tools is far past the 50+ threshold, and the count is inflated by wrapper-per-tool proliferation plus duplicate session, callback, upload, and exec capabilities. A leaner set of around 20-30 generic tool runners and shared infrastructure tools would serve the same purpose more effectively.

    Completeness4/5

    The set covers a broad pentest lifecycle: recon, scanning, exploitation, AD attacks, file transfer, pivoting, callbacks, and CTF interaction. A few niceties like vulnerability report generation or dedicated exploit execution wrappers are missing, but zebbern_exec and msf_session_execute provide practical workarounds.

    Maintenance

    ActivitySlowing
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      D
      maintenance
      A penetration testing MCP server that runs 20 hacking tools inside a Kali Linux Docker container, enabling AI assistants to execute security scans and attacks via natural language.
      2
      -
    • A
      license
      A
      quality
      B
      maintenance
      Enables AI agents to perform professional penetration testing through a containerized Kali Linux environment, exposing industry-standard offensive security tools as structured MCP tools.
      45
      6
      MIT
    • F
      license
      B
      quality
      C
      maintenance
      A Kali Linux MCP server that integrates network security and penetration testing tools like Nmap, Whois, Dig, Ping, Nikto, Hydra, and SQLMap into a Dockerized environment for use with LLMs.
      16
      3
      -