Test your MCP server in one command: conformance, sign-in, security mapped to the OWASP MCP Top 10, agent experience, what your tools cost an agent in context, and performance. A grade, plain findings, a fix for each, and a signed result anyone can verify.
npx saktum check -- npx -y @your/mcp-server # a local server
npx saktum check https://your-server.example/mcp # a remote server (sign-in checks run without a token)
Runs on your machine. Nothing is uploaded. Your own environment variables are never passed to the server.
GRADE F (trust score 49, capped: a critical check failed)
Grade F: 2 problems found (1 critical, 1 medium); 11 checks passed.
Fix first (critical): found what looks like an AWS access key in tool upload_report.
Agents spend about 142 tokens reading these tools on every turn, about $0.43 per 1,000 turns.
agent-experience 100
conformance 100
security 61
✗ FAIL CRITICAL sec.secrets.exposed No secrets in tool definitions or responses
Found what looks like an AWS access key in tool upload_report. Anyone who can list tools or trigger an error can read it.
Evidence: ess_key_id":{"type":"string","default":"AKIA…[redacted 16 chars]"}},"required":["nam
Fix: Remove keys, tokens and passwords from descriptions, schemas, defaults, examples and error messages; load them from a secret store at runtime.
✗ FAIL MEDIUM sec.errors.leakage Errors do not expose internals
The error for an unknown tool exposes a JavaScript stack trace, which helps attackers map your internals.
Evidence: at handleCall (/home/deploy/acme-mcp/src/handlers.js:42:11)
Every report ends with OWASP MCP Top 10 coverage: which categories this run tested and which come in later releases.
| Area | Checks |
|---|---|
| Conformance | Handshake, tool list, valid input schemas (draft-07, 2019-09, 2020-12), tool names, unknown-tool handling |
| Sign-in (remote servers) | HTTPS; 401 with a Bearer challenge pointing to resource metadata (RFC 9728); protected resource metadata; authorization server metadata with PKCE S256 (RFC 8414, RFC 7636); invalid tokens rejected (RFC 6750) |
| Security | Hidden instructions to the agent, invisible and bidirectional characters, terminal escape sequences, secrets in tools or errors (always redacted), stack traces and source paths in errors |
| Agent experience | Danger annotations, description quality, context cost in tokens and dollars |
| Performance (never scored) | Start-up, latency, memory of the whole process tree, and a rate-limit burst test with --active |
Validated on 31 public MCP servers before release. Each check cites the specification clause or OWASP item it verifies.
| Option | Use |
|---|---|
--env NAME=VALUE |
Give a local server what it needs to start (repeatable) |
--header "Authorization: Bearer <token>" |
Test a remote server's tools, not just its sign-in |
--html report.html |
Also write an HTML report |
--fail-on high |
Exit code for CI: fail on findings at or above this severity |
--price-per-mtok 3 |
Your model's input price, for the cost estimate |
--active |
Extra checks that send more traffic. Only against servers you own. |
npx saktum doctor -- <same target> explains why a server will not start or connect.
npx saktum verify saktum-report.json checks a signed result.
This is an early release, and feedback shapes v0.2. In two minutes:
- Something looked wrong?Report a false positive
- Something missing?Suggest a check
- What should v0.2 do?Vote and comment
Planned next: active checks for servers you own, sign-in registration and redirect checks, a lock file that catches silent tool changes, a sandbox for untrusted servers, then CI (GitHub Action, SARIF) in January.
Results are self-tested (assurance level L1): the signature proves the report was not changed after the run, not that it was run independently. Findings describe the server at the time of the run and are not a guarantee of security. Static checks can miss cleverly disguised attacks; a pass is evidence, not proof.
Apache-2.0. Security issues: see SECURITY.md.