io.github.kushvinth/crowbar
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@io.github.kushvinth/crowbar check nginx 1.18.0 for CVEs and known exploits"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
crowbar
Retrieve CVEs and index them inside your codebase.
Point it at a product and version. It pulls the CVEs that apply and indexes them locally, so any tool in your repo can query the answer.
Quick start · How crowbar works · What you can answer · Showcases · FAQ · Documentation
npx crowbar-cli setup
Get crowbar fromkushvinth/crowbar, crowbar.kushvinth.com, or the crowbar-cli package. A similar repository name or copied README does not make a download official.
A codebase with no CVE index inside it makes people guess. crowbar is the index: it retrieves NVD, CISA KEV, EPSS, SSVC, the CVE List, Exploit-DB, Nuclei, Metasploit, OSV, Debian, MSRC and Apple advisories, and writes them into a local SQLite + FTS5 database you query from your terminal, your CI or your agent. It is offline and reproducible - the same query gives the same answer, no API, no rate limits - so a lookup costs milliseconds instead of a guess.
No web UI, no daemon, no third-party Python packages. Point it at a product and version and it tells you what is actually exploitable, ranks it by real-world risk, and links the proof-of-concept.
crowbar indexes public exploit and PoC data as data. It never clones, compiles, or runs it. Treat anything it surfaces as hostile code and detonate only in a disposable VM. Use it only against systems you own or may test. See DISCLAIMER.md.
Quick start
Set up your agent
With Node.js and uv installed, run:
npx crowbar-cli setupSetup finds Claude Code, Codex, Cursor, Gemini CLI and opencode on your machine,
shows exactly which configs it will change, and applies them only after you
approve. It registers crowbar's MCP server, installs the skill, backs up every
config file it edits, and offers to build the local index. Restart your agent
afterward. --agents claude,cursor picks agents explicitly; --yes skips the
prompts. npx crowbar-cli setup --remove undoes it the same way: it lists what
it will remove, asks first, and leaves your other servers and the index alone.
Or install the Claude Code plugin
One install drops in a skill and an MCP server:
claude plugin marketplace add kushvinth/crowbar
claude plugin install crowbar@crowbarYour agent now has:
the skill - it reaches for crowbar on any "is X vulnerable / what CVEs hit this / known exploits for …" question instead of guessing from memory;
the MCP tools -
cb_cve,cb_product,cb_search,cb_exploit, … answering offline from the local index.
That's the whole setup. The first time you ask a security question, the
crowbar skill sets itself up automatically - installs the tool and builds a
small local index in ~/.crowbar (~150 MB; a few minutes, since NVD downloads are
slow), then answers. You never run a CLI command, and the index refreshes itself
once a day while you use it.
Other agents (Cursor, Codex, Gemini, opencode)
Same MCP server, one line - no heavy install, uvx runs it on demand:
{ "command": "uvx", "args": ["--from", "crowbar-cli", "crowbar-mcp"] }claude mcp add crowbar -- uvx --from crowbar-cli crowbar-mcp # or, manually, in Claude CodeThe repo ships .claude-plugin/, .codex-plugin/, gemini-extension.json and a
host-agnostic skills/crowbar/SKILL.md, so most hosts auto-detect it.
Prefer a plain CLI?
Everything the agent can do, you can do in the terminal. Python 3.11+, zero deps.
uvx crowbar-cli # run without installing
uv tool install crowbar-cli # or pipx install crowbar-cli / pip install crowbar-cli
nix run github:kushvinth/crowbar
git clone https://github.com/kushvinth/crowbar && cd crowbar && ./bin/crowbar --helpBuild an index
crowbar sync --tier lite # pick lite | medium | heavy | extreme
crowbar build --tier lite # build the local index
crowbar doctor # verify it's readyWith uvx, prefix each: uvx crowbar-cli sync --tier lite. To pull a
prebuilt mirror instead of the upstream feeds, set CROWBAR_MIRROR - see
docs/DATASETS.md. Refresh later with
crowbar update (~10 s) - NVD modified + KEV + EPSS + Exploit-DB, in place.
Size tiers - each a superset of the one above, so you pick your disk budget:
tier | size | includes |
| ~150 MB | KEV, EPSS, recent CVEs, CWE names |
| ~500 MB | + full NVD history, CAPEC, SSVC, ATT&CK |
| ~2 GB | + Exploit-DB, Nuclei, Metasploit, OSV, PoC |
| full | + MSRC, Apple, Debian - every source |
Related MCP server: NVD/NIST MCP server
Fresh, and fast
It keeps itself current. Each CLI call and MCP session checks the index age. Once it is a day old, crowbar refreshes it in the background: NVD changes, KEV, EPSS for every CVE, new PoCs and CVE chatter. An index past NVD's 8-day change window gets a re-sync and rebuild, so no gap is left behind. A refresh that fails (you were offline, or an upstream was down) never stamps the index fresh.
crowbar doctor,cb_statsand the MCP answers themselves report how old the data is and why the last refresh failed.It knows what's moving.
crowbar trendingranks CVEs by how many places name them: Hacker News, oss-security, Full Disclosure, ZDI, Mastodon, and ENISA's exploited list (the EU counterpart to CISA KEV).crowbar newshows KEV/EUVD additions, new CVEs and new PoCs since a date.crowbar cvelists where each CVE was discussed.Speed you can measure.
crowbar benchtimes the queries agents run most on your index. They are local SQLite lookups on indexed columns, with no network, API key or rate limit.
How crowbar works
Your agent asks crowbar about a target over MCP. crowbar answers from the local index - never the network - with the KEV/SSVC/EPSS/CVSS verdict, the affected version ranges, and links to working PoCs. The agent uses that grounded result to explain the risk, decide what to patch, or write and test an exploit under authorization. CLI commands use the same read layer.
What you can answer
Ask | What crowbar returns | Command |
What does this CVE mean, and is it serious? | CVSS, EPSS, KEV, SSVC, CWE, CPE ranges, refs, PoCs |
|
We run product X version Y - what's exploitable? | Only CVEs whose CPE range covers Y, with the matched range |
|
Is it exploited in the wild? | CISA KEV + SSVC |
|
What is everyone talking about this week? | CVEs ranked by how many places post about them |
|
What changed since Monday? | KEV/EUVD exploited additions, new CVEs, new PoCs |
|
Where's a working exploit or detection? | Exploit-DB, Nuclei templates, Metasploit modules |
|
A public PoC repo for this CVE? | GitHub PoC metadata (stars, language) - never cloned |
|
Is this npm / PyPI / Go package affected? | OSV/GHSA advisories with introduced/fixed ranges |
|
Windows / iOS / macOS advisory? | MSRC + Apple advisories, normalized impact (RCE/EoP/…) |
|
The exploit code, not just a title? | Full-text search inside 46k Exploit-DB sources |
|
A living-off-the-land binary? | GTFOBins / LOLBAS / PayloadsAllTheThings |
|
A technique, weakness class, or model? | ATT&CK / CWE / CAPEC / LLM catalog |
|
Anything, my way | read-only SQL over the whole index |
|
Rank threat by KEV / SSVC active > EPSS > CVSS - never CVSS alone. Most read
commands accept --json. See the full command reference
and AGENTS.md for the traversal manual.
Showcases
Triage Log4Shell in one command. crowbar cve CVE-2021-44228 returns CRIT 10.0 · KEV · EPSS 1.00 · ssvc=active, 44 affected CPE ranges, Exploit-DB 50592
(RCE), 44 Nuclei templates and 4 Metasploit modules - the whole verdict, offline,
in ~80 ms.
"What do we run that's exploitable?" crowbar product nginx --version 1.18.0
tests every NVD CPE range against 1.18.0 and returns only real hits, like
CVE-2021-23017 HIGH 7.7 · matched f5:nginx >=0.6.18 <1.20.1 - not every CVE that
merely mentions "nginx".
Hunt the actively-exploited. crowbar search "remote code execution" --kev --epss 0.5 --sort epss surfaces CVE-2021-26855 (Exchange), CVE-2020-0796
(SMBGhost) and friends, each with exp×N nuc×N poc×N counts, ranked by real
exploitation probability.
Used crowbar on something interesting? Open an issue or PR with the target, your question, and what you found.
FAQ
Any MCP host: Claude Code, Codex, Cursor, Gemini CLI, opencode, and Claude
Desktop. The launch command is uvx --from crowbar-cli crowbar-mcp; this repo
ships ready manifests for the common ones. You can also use every capability
from the terminal.
No. Every dataset is fetched over plain HTTP at build time with no token; after that crowbar is fully offline. The MCP server makes zero network calls, and the index opens read-only.
Your choice. A lite index is ~150 MB and answers in milliseconds; extreme is
the full corpus (~1.5 GB index over ~3.3 GB of raw data). crowbar doctor tells
you what you built and crowbar stats is always authoritative.
Never. It indexes exploit and PoC data as data. crowbar exploit <id>
downloads source for you to inspect; running it is your decision, in a disposable
VM, under authorization.
Those query live APIs (network, rate limits, answers that drift). crowbar is a local index: offline, reproducible, and it full-text-searches Exploit-DB source code, which the API wrappers don't.
Documentation
crowbar.kushvinth.com - quick start, commands, agent setup.
AGENTS.md - the agent traversal manual (schema, recipes, guardrails).
docs/DATASETS.md - per-source shape, license and caveats.
docs/MAP.md - repo tree, data flow, table map.
Star history
Disclaimer
crowbar is for lawful security research, defense, and analysis of systems you own or are authorized to test. See DISCLAIMER.md. KEV/SSVC describe exploitation observed elsewhere, EPSS is a model estimate, and a missing row is not proof of "not affected" - verify before you act.
Contributing
We'd love your help with crowbar! Open an issue to report a bug, a wrong CVE record or a missing dataset, or send a pull request to improve the code or docs.
See CONTRIBUTING.md for the dev loop, the add-a-dataset checklist and the test suites. crowbar is standard-library only; keep it that way.
License
crowbar-authored code is licensed under MIT. The upstream data it indexes keeps its own licenses and terms; see Third-party notices and docs/DATASETS.md for the source inventory.
Indexed exploit and PoC data is reference material, not permission. crowbar never runs it, and nothing it returns authorizes testing a system. Use it only against systems you own or are allowed to test; see DISCLAIMER.md.
Available Tools
22 toolscb_advisoryB
OSV/GHSA advisories with introduced/fixed version ranges. Filter by ecosystem (npm, PyPI, Go, Maven, NuGet, RubyGems, crates.io, Packagist, Debian, Alpine) or exact package name.
Inputs: limit, query, package, ecosystem.
cb_attackC
MITRE ATT&CK technique lookup (enterprise, mobile, ics).
Inputs: query (required), limit, domain.
cb_binaryC
Living-off-the-land binary abuse: GTFOBins (unix), LOLBAS (windows), PayloadsAllTheThings (web payloads).
Inputs: query (required), limit, source.
cb_codeB
Full-text search inside Exploit-DB source code; returns path + snippet.
Inputs: query (required), limit.
cb_cveB
One CVE in full: CVSS, EPSS, KEV, CWE, CPE version ranges, references, Exploit-DB / Nuclei / Metasploit hits, SSVC decision points.
Inputs: id (required).
cb_cweC
CWE definitions (id or keyword).
Inputs: limit, query.
cb_debianC
Debian security tracker: per-release status and fixed version for a package.
Inputs: package (required), cve, limit.
cb_exploitB
Search the Exploit-DB index (46k entries), offline. To download the source for a numeric id, use the CLI: crowbar exploit <id>. Nothing is executed.
Inputs: query (required), limit.
cb_kevC
CISA Known Exploited Vulnerabilities catalog.
Inputs: limit, query.
cb_modelsC
LLM catalog: context window, price per M tokens, capabilities.
Inputs: limit, query, provider.
cb_msfC
Search Metasploit module paths.
Inputs: query (required), limit.
cb_newA
What changed since a date (default 7 days): CISA KEV and ENISA exploited additions, newly published CVEs (KEV > EPSS > CVSS first), new public PoCs. Each part has total and the top rows.
Inputs: days, limit, since.
cb_nucleiC
Search Nuclei templates; returns YAML paths for exact matchers.
Inputs: query (required), limit, domain.
cb_osB
Vendor OS advisories: Windows (MSRC Security Update Guide) and Apple iOS/iPadOS/macOS security releases. Filter by platform, impact (RCE, EoP, DoS, InfoDisclosure, Spoofing, Bypass) or exploited.
Inputs: limit, query, since, impact, +2 more.
cb_pocA
Public GitHub PoC repositories (PoC-in-GitHub metadata). Never executed. Filter by CVE, language or minimum stars, or list the CVEs with the most PoCs.
Inputs: cve, top, limit, query, +2 more.
cb_productA
CVEs affecting a product/vendor. Pass version to test NVD CPE version ranges, e.g. {name: nginx, version: 1.18.0}.
Inputs: name (required), limit, version.
cb_searchC
Full-text search across every CVE (2002-today). Ranked by relevance blended with CISA KEV, EPSS and public-exploit availability. Filters: kev, exploit, min_epss, min_cvss, severity, year, since, cwe, product, vendor, ssvc, poc.
Inputs: cwe, kev, poc, sort, +12 more.
cb_sourcesA
Source provenance: which upstream source filled the index, its URL, license, SHA-256, load time and row count.
Inputs: no named parameters.
cb_sqlC
Read-only SQL. Tables: cve, cve_cwe, cve_product, cpe_match, cve_ref, cve_exploit, cve_nuclei, cve_msf, kev, exploit, shellcode, nuclei, msf, attack, cwe, capec, model, ssvc, advisory, package_range, debian, os_advisory, toolkit, cve_cna, cve_meta, poc, poc_orphan, cve_alias, source_state, exploit_code_fts, toolkit_fts, meta.
Inputs: query (required), limit.
cb_statsA
Corpus health and freshness. index.age_days is days since the last update; index.stale means older than the 8-day NVD change window, so recent CVEs may be missing or outdated - say so and run crowbar update. index.last_error is why the last background update failed.
Inputs: no named parameters.
cb_tablesA
Data map: every index table, what it holds, and which command/dataset feeds it.
Inputs: no named parameters.
cb_trendingB
CVEs people are posting about right now, ranked by how many distinct places name them (Hacker News, oss-security, Full Disclosure, ZDI, Mastodon, ENISA EUVD exploited) with KEV/EPSS/CVSS beside each.
Inputs: days, limit.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 22 tools
Each tool maps to a clearly named resource (CVE, product, KEV, Exploit-DB, Nuclei, MSF, ATT&CK, CWE, advisories, PoCs, Debian, binaries), so misselection is unlikely for the core lookups. Minor blur exists between cb_exploit and cb_code (both over the Exploit-DB index), between cb_search and cb_product (search also accepts product/vendor filters), and among the four meta tools cb_sql, cb_tables, cb_sources and cb_stats, though their descriptions do differentiate them.
Every tool uses the same cb_ prefix with a single short token, giving a very predictable namespace. The tokens mix nouns (cb_cve, cb_kev, cb_msf) with verbs/adjectives (cb_search, cb_new, cb_trending), which is a minor stylistic deviation but remains immediately readable and consistent in form.
22 tools is on the heavy side, but the server aggregates roughly a dozen distinct upstream sources and most get exactly one dedicated entry point. The only soft bloat is the four introspection/meta tools (sql, tables, sources, stats), which could arguably be consolidated to two.
The surface covers discovery (search, trending, new, product), per-CVE detail, exploitation artifacts (Exploit-DB, code, PoCs, Nuclei, MSF, LOLBAS/GTFOBins), enrichment (KEV, EPSS via filters, SSVC, CVSS, CWE, ATT&CK), ecosystem advisories (OSV/GHSA, Debian, vendor OS), plus data provenance, schema map and a read-only SQL escape hatch for anything not wrapped. Gaps are trivial: CAPEC and shellcode tables are only reachable through cb_sql, and cache refresh must be done via CLI, both workaroundable.
Maintenance
Related MCP Connectors
ZEN SecDB MCP server for CVE intelligence, CVSS/EPSS scoring, advisories, SSVC, and package audits.
Read-only CVE intelligence, remediation playbooks, and agent setup guides. Not a scanner.
Read-only vulnerability intelligence: CVEs, exploitation evidence, fixes and product matches.
Real-time CVE, exploit, and vulnerability intelligence for AI assistants (350K+ CVEs, 115K+ PoCs)
Related MCP Servers
- AlicenseAqualityDmaintenanceEnables querying and analyzing a legacy, file-based vulnerability registry via MCP tools for CVE lookup, search, vendor navigation, and aggregate statistics.5ISC
- AlicenseAqualityCmaintenanceEnables users to search and retrieve CVE records, CVE history, CPE Dictionary entries, and CPE match criteria through MCP tools, with local caching, pagination, and upstream rate-limit handling.10MIT
- FlicenseNot gradedqualityAmaintenanceEnables querying the CVE-Search API to browse vendors and their products, retrieve vulnerabilities per vendor/product, look up specific CVE IDs, fetch the latest 30 CVEs with CAPEC, CWE and CPE expansions, and check database metadata. This gives MCP clients conversational access to public vulnerability data.-
- FlicenseNot gradedqualityAmaintenanceEnables querying the CVE-Search API to browse vendors and products, retrieve vulnerabilities by vendor/product or CVE ID, fetch the latest CVEs with CAPEC/CWE/CPE details, and check database update status via MCP tools.-