mcp-posture

Security posture scanner for remote MCP servers. Point it at a Streamable HTTP endpoint and it audits the OAuth setup (MCP Authorization spec, RFC 9728, RFC 8414, PKCE, RFC 8707, Client ID Metadata Documents), transport hardening and the tool surface (poisoning, shadowing, rug pulls), with spec-cited findings and CI-native output (SARIF, JSON, Markdown).

Only scan servers you own or are authorized to test.

Why

Most MCP scanners inspect local client configs and stdio servers. mcp-posture focuses on remote servers and what an attacker sees from the outside before logging in: how the server challenges, which authorization server it trusts, whether that server enforces PKCE S256, whether tokens are audience-bound, whether the tool descriptions hide instructions. Every finding cites the spec section and the MCP revision it applies to (2025-03-26 through 2026-07-28).

Quickstart

# no install needed (or: pipx install mcp-posture / uv tool install mcp-posture)
uvx mcp-posture scan https://mcp.example.com/mcp

# machine-readable outputs, fail the build on high or worse
mcp-posture scan https://mcp.example.com/mcp --sarif results.sarif --markdown summary.md --fail-on high

# lint your own client's Client ID Metadata Document
mcp-posture cimd lint https://app.example.com/oauth/client.json

# scan every remote server declared in your MCP clients (.mcp.json, Claude, VS Code, Cursor, Windsurf)
mcp-posture discover
mcp-posture scan --from-client-config auto

# list tools behind authentication: log in to your own server in the browser (token never shown)
mcp-posture scan https://mcp.example.com/mcp --login
# ... or pass a token through the environment, never on the command line
MCP_TOKEN=... mcp-posture scan https://mcp.example.com/mcp --token-env MCP_TOKEN

Docker (distroless, nonroot, signed with cosign):

docker run --rm ghcr.io/batou9150/mcp-posture scan https://mcp.example.com/mcp

GitHub Action (SARIF to code scanning, Markdown to the job summary):

- uses: batou9150/mcp-posture@main   # pin a tag or SHA
  with:
    targets-file: mcp-servers.txt
    fail-on: high

Claude Code skill: the scanner plus a semantic review of tool descriptions, prioritization and remediation snippets for common authorization servers and gateways.

/plugin marketplace add batou9150/mcp-posture
/plugin install mcp-posture@mcp-posture

Then ask "audit the security of https://mcp.example.com/mcp".

Sample output (local misconfigured fixture, scripts/demo_servers.py):

http://127.0.0.1:8402/mcp
  spec 2025-06-18 (negotiated) · transport streamable-http · auth not required
critical MCPP-ASM04   code_challenge_methods_supported ['plain'] does not include S256.
high     MCPP-ASM02   Metadata declares issuer 'http://127.0.0.1:8402/', expected 'http://127.0.0.1:8402'.
high     MCPP-AUTHN01 tools/list answered without credentials (2 tool(s) including delete_note).
high     MCPP-PRM06   bearer_methods_supported includes `query`.
high     MCPP-TRN08   Session IDs look sequential (numerically close across two sessions).
medium   MCPP-PRM03   resource '.../mcp/' does not match '.../mcp' (trailing slash only).
...

What it checks

mcp-posture checks list prints the catalogue; mcp-posture checks show MCPP-ASM04 explains one check (rationale, remediation, references). Check IDs are stable and never reused. The research behind the catalogue, with every MUST/SHOULD mapped to a check, is in docs/spec-notes.md.

The default mode is passive: metadata GETs plus the standard MCP handshake and list calls. No tool is ever called. Behaviours that only an active probe could confirm (Origin validation, token audience enforcement, PKCE enforcement) are listed as not verified in docs/spec-notes.md.

How it works

flowchart LR
  T[Targets<br/>URLs · targets file · client configs] --> C[Collect<br/>hardened HTTP client]
  C --> P[MCP prober<br/>server/discover → initialize → SSE]
  C --> D[OAuth discovery<br/>401 challenge · PRM · AS metadata · TLS]
  P & D --> X[Immutable scan context]
  X --> K[Checks<br/>pure functions, registry]
  K --> S[Suppressions · baseline]
  S --> R[Reports<br/>table · JSON · SARIF · Markdown]

Each target is collected once (all network I/O, through a client that refuses private addresses at connect time), frozen into a context, then every check runs as a pure function over it. A failing check becomes an MCPP-ERR00 finding instead of aborting the scan.

CI usage

# mcp-posture.toml (CLI flags override it)
[scan]
targets_file = "mcp-servers.txt"
fail_on = "high"
baseline = "mcp-posture.lock.json"   # rug-pull detection, created with `mcp-posture pin`
disable = ["ASM08"]

# .mcp-posture-ignore: every suppression needs a justification; expired ones resurface
[[ignore]]
check = "MCPP-ASM09"
target = "https://mcp.example.com/*"
justification = "Open DCR is intended: public client registry"
expires = 2026-12-31

SARIF results are anchored to the line of the file that declares each target (targets file, mcp-posture.toml, .mcp.json), so GitHub code scanning can display them.

Safety of the scanner

• Tokens are read only from an env var, a file or stdin, and are redacted from every output and log.
• Private, loopback, link-local and cloud-metadata addresses are refused at connect time (after DNS resolution, every redirect hop included) unless --allow-private is passed.
• Credentials are never forwarded across origins; response size and time are capped.
• Server-controlled text is neutralized in reports: no raw control, bidi or invisible characters reach your terminal or PR comments.
• No telemetry.

Compared with other MCP scanners

Several good tools exist; they solve different problems. Facts as of 2026-10 (see docs/spec-notes.md for the survey):

Pair mcp-posture with a runtime proxy (e.g. mcp-context-protector) for enforcement; it reports, it does not block.

Responsible use

Scan only servers you own or have written permission to test. Passive mode sends a handful of standard requests; even so, unsolicited scanning of third-party infrastructure may violate their terms or the law.

Contributing and security

See CONTRIBUTING.md (how to add a check) and SECURITY.md (reporting vulnerabilities, verifying release signatures and attestations).

Development

uv sync
uv run ruff check && uv run ruff format --check && uv run mypy && uv run pytest
uv run python scripts/demo_servers.py   # local fixtures to scan with --allow-private
uv run --group docs mkdocs serve         # docs site

License

MIT © Baptiste PIRAULT