litmus

The open behavioral litmus harness for MCP servers — grade A–F, reproducible.

Grade a server in one command

# -p is required: the package ships three bins, so npx must be told which to run
npx -y -p @polygraphso/litmus polygraphso-litmus litmus npm/@modelcontextprotocol/server-filesystem

Point it at an npm ref, a pypi ref, a github/owner/repo ref (cloned, built, and run sandboxed — Docker required; the grade pins the commit SHA), an https:// MCP endpoint, or a local entry file. The harness connects the way an agent would, fingerprints the exact tool surface, runs the four probe categories, and prints the grade with the findings behind it — plus a deterministic evidence bundle on disk. It runs the target's code (egress is Docker-sandboxed; without Docker, C-02 is skipped and the grade caps at B), takes ~20–60s, and exits non-zero on D/F so it scripts anywhere. To dispute any published grade, re-run this same command against the same server — open and deterministic means a re-run reproduces the grade, or refutes it.

Looking up a grade someone already published takes under a second and runs nothing: the polygraph.so index, or check_server from the MCP tools below.

Use it from your agent — MCP server + plugin

For grade lookups, point any MCP client at polygraph's hosted endpoint, no install:

claude mcp add --transport http polygraph https://polygraph.so/api/mcp

or the raw config:

{ "mcpServers": { "polygraph": { "url": "https://polygraph.so/api/mcp" } } }

This serves the lookup tools only (check_server, list_servers, request_grade); grading a server yourself (run_litmus, run_skill_litmus) executes its code, so it needs the local stdio install below.

The package also ships a stdio MCP server (polygraphso-litmus-mcp) with the full toolset, for any MCP-capable client:

• check_server — read a server's published grade in under a second (no execution); the pre-flight check before recommending or installing a server.
• list_servers — servers with a published grade, A first; paged (default 25 per call, with grade/limit/offset filters and a full-corpus summary).
• request_grade — record a grade request with polygraph.so ($1 one-time fee; graded within 48h of payment — the response carries the payment link).
• run_litmus — grade a server now: the full harness, grade + evidence returned to the agent.
• run_skill_litmus — grade a Claude Code / Agent Skill (static scan, A/B/D/F).
• verify_attestation — read the onchain proof behind a published grade (EAS on Base).

In Claude Code, the plugin wires the server plus two commands in one step:

/plugin marketplace add polygraphso/litmus
/plugin install polygraph@polygraphso

then /polygraph:grade <server> and /polygraph:check <server>. Cursor and manual JSON setups are on polygraph.so; full tool docs in packages/litmus/README.md.

Gate your CI on MCP grades — GitHub Action

Fail a build when an MCP server or an Agent Skill it ships grades D/F under the open behavioral litmus. For servers it is hybrid — a fast lookup of the published grade, then the harness when ungraded; for skills it is a fast static scan. Un-gradeable targets warn unless strict.

It's on the GitHub Marketplace as polygraphso/litmus@v1. For a security gate, pin to a commit SHA rather than the mutable @v1 tag:

# .github/workflows/mcp-gate.yml
name: mcp-gate
on: [pull_request]            # NOT pull_request_target — that exposes secrets to fork PRs
permissions:
  contents: read
jobs:
  gate:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v5
      - uses: polygraphso/litmus@<commit-sha>   # pin to a SHA; resolve from the v1 release
        with:
          # Name the targets explicitly (recommended). Grading runs a server's code,
          # so on a public repo prefer an allowlist over discovering PR-controlled config:
          servers: |
            npm/@modelcontextprotocol/server-filesystem
          skills: |
            ./my-skill
          # discover: "true"  # opt in to auto-discovery (.mcp.json/.vscode/.cursor) — trusted repos only
          # min-grade: B      # stricter than the default D/F gate
          # strict: "true"    # also fail on targets that cannot be graded

Inputs: servers · skills · discover (default false) · min-grade · strict · working-directory · version · bearer. Outputs: result · failed · report.

Security. Grading a server runs its code (egress is Docker-sandboxed, but it still executes). Trigger on pull_request, never pull_request_target. Keep discover off on public repos and name targets explicitly — auto-discovered config is pull-request-controllable. bearer is sent as an Authorization header to the target, so pass it only for an explicitly trusted, pinned remote — never with discovery or on untrusted PRs, and keep it scoped and short-lived.

Not on GitHub? The gate is a plain command — npx @polygraphso/litmus@0.20.0 ci (pin the version) — so it runs in any CI or as a pre-commit hook. A grade is a measurement, not a guarantee: re-run the open harness to reproduce any result.

What litmus is

This is the source for @polygraphso/litmus, the open behavioral litmus harness for MCP servers from polygraph.so.

The harness connects to an MCP server the way an agent would, fingerprints its exact tool surface, and runs four probe categories — C-01 tool-output injection (static, dynamic, and second-order — one tool's output weaponized as another's input), C-02 permission/egress (in a hardened default-deny Docker sandbox, matched host and port), C-03 sensitive-data handling (planted canaries), C-04 adversarial-input handling (malformed/oversized and jailbreak inputs) — then grades the server A–F. A passing grade is a measurement, not a guarantee; the methodology and its disclosed limits are at polygraph.so (the open source here is the ground truth).

Alongside the grade, an npm target's dependency tree is checked against the osv.dev vulnerability database and any vulnerable dependencies are reported as dependency advisories. This is a separate, point-in-time signal — it is advisory only: it never affects the A–F grade and is not part of the reproducible evidence (vulnerability data changes over time, so folding it into the grade would break re-run reproducibility). It applies to npm targets only; other target kinds report it as skipped. Resolution runs npm install --package-lock-only --ignore-scripts, which resolves the tree without downloading tarballs or running any package code. Opt out with --no-deps-audit (or LITMUS_DEPS_AUDIT=0).

The same package also grades Claude Code / Agent Skills (a SKILL.md + bundle) under a separate static litmus (litmus-skill-v3): a deterministic byte-scan — S-01 prompt injection, S-03 data-exfiltration instructions, S-04 dangerous commands in the SKILL.md body or bundled scripts (incl. base64-obfuscated curl | bash) — graded A/B/D/F and anchored by a whole-directory content hash, plus a separate advisory quality signal. It is static (no execution): an A is static-clean, not behavioral proof. See packages/litmus/README.md.

The hosted, operator-run grading service is not in this repo — it lives in a separate private repo and consumes this package from npm like any other client.

Layout

This is a pnpm monorepo. Only @polygraphso/litmus is published; the @polygraph/* packages are private building blocks that tsup bundles into it.