provekit-mcp

A hardened MCP server that gives an AI agent a code security scanner, built so the tools themselves cannot be turned against the host. It ships with its own red-team suite that spawns the real server and attacks it over the protocol, and proves every trust boundary holds.

An MCP tool is a function a model can call with arguments it chose, sometimes while reading untrusted content. So the interesting question about an MCP server is not "what can it do" but "what happens when someone points it at ../../../../etc/passwd." This one is built to answer that question out loud.

python -m provekit_mcp.server        # run the server (stdio)
python -m redteam.run                # attack it and print the verdict
pytest -q                            # 97 tests, incl. the live red-team

---

Two tools, both built for a hostile caller

Tool	What it does	Why it's safe
scan_code(code, filename)	Scans a snippet the agent already has in hand for leaked secrets and insecure patterns (OWASP Top 10).	No filesystem access at all, so there is no path to traverse. filename is reduced to a bare label and never read. Input size is bounded.
scan_path(path)	Scans a source file, but only inside a configured workspace root.	Every path goes through guard.safe_resolve first: absolute paths, .. traversal, null bytes, and symlink escapes are all refused before a single byte is read.

The threat model (this is the point)

An MCP server hands an autonomous model a set of tools. The model may be acting on content it just fetched from the web, an issue comment, a file it read, any of which can carry an instruction the model wasn't supposed to follow. So the server has to assume every argument is attacker-controlled. The trust boundaries provekit-mcp defends:

1. Path confinement. A file tool must never read outside the workspace it was given. The classic breakouts, ../ traversal, an absolute /etc/passwd, a null byte to truncate an extension check, and a symlink inside the root pointing out of it, are each closed and each has a test.
2. Resource bounds. A single tool call must not be a way to exhaust host memory or CPU. Per-call input is capped; oversized calls are refused in milliseconds, not after allocating.
3. No catastrophic backtracking. The scanner's rules are all bounded regexes. A 40,000-character pathological line scans in single-digit milliseconds, so a crafted argument can't hang the server (ReDoS).
4. Arguments are inert data. A malicious string passed as code is scanned as text, never executed. The scanner reads it, flags the eval / os.system in it, and moves on.
5. Refusals don't leak. A rejected call returns a structured { "ok": false, "code": "escape", ... }, never a stack trace, never a partial read.

Defense in depth: guard.safe_resolve is the primary path control, and the MCP SDK's own ResourceSecurity(reject_path_traversal, reject_absolute_paths, reject_null_bytes) is enabled as an independent second layer. Neither is trusted to be the only thing standing between a tool call and the filesystem.

The red-team suite

python -m redteam.run doesn't test the functions in-process, it spawns the actual server as a subprocess and speaks MCP to it (initialize → tools/list → tools/call), firing each attack the way a hostile client would. Every response is triaged into one of five honest outcomes:

• HELD — an attack was correctly refused
• BREACH — an attack succeeded (critical)
• OK — a legitimate call worked
• REGRESSION — a legitimate call was wrongly refused (over-blocking is a real failure; a scanner nobody can use is worthless)
• INCONCLUSIVE — no usable answer

That last outcome is the discipline that matters. A cold start, a hang, or a garbled frame is never scored as "secure." The run is only clean when there are zero breaches, zero regressions, and zero inconclusive results, every control actually verified, not assumed.

attacks held: 9   breaches: 0   controls OK: 3   regressions: 0   inconclusive: 0
VERDICT: ALL CONTROLS HELD AND VERIFIED

Attacks currently in the suite: path traversal, deep traversal, absolute path, symlink escape, null-byte truncation, binary-file read, 6 MB resource exhaustion, a ReDoS pathological line, and an argument-as-code injection, plus two control calls that prove the guards don't over-block. The machine-readable result is written to assets/redteam-report.json.

What the scanner catches

The engine is a Python port of provekit, kept rule-for-rule compatible on the shared detectors and extended with Python-specific vulnerabilities (since that's where the work is).

OWASP	Examples
A07 / A02 — Leaked secrets	AWS / GitHub / Stripe / OpenAI / Anthropic keys, private-key blocks, DB URLs with inline credentials, hard-coded passwords
A03 — Injection	eval / new Function, os.system built from an f-string, shell commands via interpolation, subprocess(..., shell=True), SQL by concatenation, innerHTML
A08 — Insecure deserialization	pickle.loads, yaml.load without SafeLoader
A02 — Broken crypto / transport	requests(..., verify=False), rejectUnauthorized: false, MD5/SHA1 for passwords, Math.random()/random for tokens
A10 — SSRF	user-controlled input reaching a server-side HTTP request
A05 — Misconfiguration	wildcard CORS, debug=True

It is built to be precise, because a scanner that cries wolf is a scanner you switch off. It skips parameterized SQL, env-var reads, bcrypt/argon hashes, yaml.safe_load, and placeholder values; it stays quiet in test/ and fixture files on the insecure things test code does on purpose, while still catching a real key anywhere. And it never silently skips a long line, a secret hidden behind a wall of padding is still caught, and a line genuinely too long to scan safely is reported (line-too-long), never dropped.

Install and wire it into Claude

git clone https://github.com/Th3Circle-app/provekit-mcp && cd provekit-mcp
python -m venv .venv && source .venv/bin/activate
pip install .          # installs the `provekit-mcp` entrypoint

For development, you can also run it straight from the repo without installing:

pip install mcp>=2.0
python -m provekit_mcp.server        # run from the repo root
pytest -q                            # 97 tests, incl. the live red-team

Add it to Claude Desktop / Claude Code (claude_desktop_config.json), pointing the workspace root at the repo you want scannable, see claude_desktop_config.example.json:

{
  "mcpServers": {
    "provekit": {
      "command": "python",
      "args": ["-m", "provekit_mcp.server"],
      "env": { "PROVEKIT_MCP_ROOT": "/absolute/path/to/your/repo" }
    }
  }
}

Now your agent can call scan_code before it ships a snippet, or scan_path to check a file, and the server guarantees it can only ever read inside that one root.

Layout

provekit_mcp/
  scanner.py   # the detection engine: bounded rules, no silent skips, ReDoS-safe
  guard.py     # the trust-boundary guards: safe_resolve, size + binary limits
  server.py    # the MCP server; tool logic lives in plain functions the tests call
redteam/
  engine.py    # HELD / BREACH / OK / REGRESSION / INCONCLUSIVE triage
  run.py       # spawns the real server over stdio and attacks it
tests/         # 97 tests: scanner correctness, the guards, the tools, the live red-team

Design notes worth reading the code for

• The tool logic is not inside the @app.tool decorators. do_scan_code / do_scan_path are plain module functions; the MCP wrappers are three lines each. This means the tests and the red-team exercise exactly what ships over the wire, not a parallel copy.
• safe_resolve uses realpath + a trailing-separator containment check. realpath collapses .. and follows symlinks, so a link out of the tree resolves to its true location and fails containment. The trailing os.sep on the prefix check prevents the /a/b vs /a/bc false pass.
• Inconclusive ≠ secure. Carried over from redteam-loop: the earlier version once scored a cold-start HTTP None as a pass. It doesn't anymore, here or there.

Who's behind it

Built by Harrison C. Songolo. Companion projects: provekit (the scanner as a zero-dep CLI + CI gate), redteam-loop (attack → propose fix → re-fire the exact exploit to prove it's closed), and security-assessments (SSRFs found, fixed, and disclosed in open-source tools).

MIT.