Click on "Install 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., "@provekit-mcp Scan this code for leaked secrets and insecure patterns."
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.
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-teamTool | What it does | Why it's safe |
| 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. |
| Scans a source file, but only inside a configured workspace root. | Every path goes through |
Related MCP server: agentguard
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:
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.
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.
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).
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.
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.
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 VERIFIEDAttacks 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.
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 |
|
A08 — Insecure deserialization |
|
A02 — Broken crypto / transport |
|
A10 — SSRF | user-controlled input reaching a server-side HTTP request |
A05 — Misconfiguration | wildcard CORS, |
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.
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` entrypointFor 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-teamAdd 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.
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-teamThe 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.
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.
Scan a snippet of code for leaked secrets and insecure patterns
(OWASP Top 10). Returns findings sorted by severity. No filesystem
access. filename is an optional label used for reporting.
| Name | Required | Description | Default |
|---|---|---|---|
| code | Yes | ||
| filename | No | snippet.txt |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description carries the burden of disclosure. It discloses a key behavioral trait: 'No filesystem access', and explains the output behavior ('Returns findings sorted by severity') and the role of the filename parameter. It does not mention potential side effects, but for a scan tool, the absence of filesystem access implies a safe, read-only operation.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is three concise sentences, each earning its place: purpose, output, and key limitation/parameter clarification. It is front-loaded with the main verb and resource, and contains no unnecessary words or repetition.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a simple two-parameter tool with no output schema and no annotations, the description provides the core information: what it does, what it returns (findings sorted by severity), and a key limitation (no filesystem access). The missing detail is the structure of findings, but the description is otherwise sufficient for an agent to invoke the tool correctly.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The schema has 0% description coverage, so the description must compensate. It adds meaning to 'filename' explicitly as 'an optional label used for reporting', but for 'code' it relies on the overall tool description ('Scan a snippet of code') to infer its purpose. This partial compensation is helpful but incomplete, especially for the required 'code' parameter.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool's action and scope: 'Scan a snippet of code for leaked secrets and insecure patterns (OWASP Top 10).' It also mentions returning findings sorted by severity, giving a complete picture. This distinguishes it from sibling tools by focusing on 'snippet' and explicitly stating 'No filesystem access', which contrasts with scan_path.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description implies usage for code snippets and excludes filesystem scanning via 'No filesystem access', but it does not explicitly name alternatives or say 'use scan_path for filesystem scanning'. The guidance is implicit rather than explicit, so it does not fully meet the 'when-to-use vs alternatives' criterion.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Scan a single source file for leaked secrets and insecure patterns.
path must be relative to the server's workspace root; paths that
escape the workspace are refused.
| Name | Required | Description | Default |
|---|---|---|---|
| path | Yes |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries full burden. It discloses that paths must be relative to the workspace root and that escaping paths are refused, which is valuable. However, it does not explicitly state that the operation is read-only, nor describe the output or error behavior beyond path rejection.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is compact and front-loaded with the main action, followed by a critical constraint. Every sentence earns its place with no redundancy or filler.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a simple single-file scan tool with one parameter and no output schema, the description covers the essential purpose and the key path constraint. It could additionally mention typical output format or error cases, but overall it is sufficiently complete for an agent to invoke correctly.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The schema only lists 'path' as a string, but the description enriches it by specifying it must be relative to the workspace root and that escaping paths are refused. This gives the parameter meaning beyond the raw schema, compensating for the 0% schema description coverage.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool scans a single source file for leaked secrets and insecure patterns. The verb 'scan' and specific resource 'single source file' make the purpose distinct, especially against the sibling scan_code.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description implies usage for scanning a specific file, but does not explicitly compare to scan_code or state when to prefer one over the other. There are no exclusions beyond the path constraint, which is more a security limitation than a usage guideline.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
The two tools are clearly distinguished by their input type: scan_code takes a code snippet, while scan_path takes a file path. Though both perform similar security scanning, the descriptions make the boundary unambiguous, so an agent should rarely misselect.
Both tools follow a consistent 'scan_<target>' pattern with snake_case, making the naming predictable and intuitive. The verb-noun structure is uniform and there are no stylistic deviations.
With only two tools, the server feels thin for a security scanning purpose. The count is borderline—not egregiously small, but it offers only the most basic scanning operations and lacks the breadth one might expect.
The server only supports scanning a snippet or a single file, with no ability to scan directories, repositories, or multiple files at once. This is a significant gap for a security scanner, as users would likely need batch or recursive scanning to be practical.
Unclaimed servers have limited discoverability.
If you are the server author, to access and configure the admin panel.
Security firewall for AI agents — scans MCP calls for injection, secrets, and risks.
Zero-config MCP security scanner for AI-generated apps. 25K+ vulnerability patterns.
Scan configs, files, or text for leaked secrets and obvious misconfigurations. Nothing stored.
Deep security scans of repos you own from your editor: dependency CVEs, SAST, git-history secrets.
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/Th3Circle-app/provekit-mcp'
If you have feedback or need assistance with the MCP directory API, please join our Discord server
