Evidence Signal MCP

The fastest way to run Evidence Signal. No connector setup — Claude runs the skill locally and your ticket content never leaves your machine.

1. In Claude, go to Settings → Capabilities and enable Code execution and file creation.
2. Go to Customize → Skills → “+” → Create skill and upload the Evidence Signal skill package.
3. Toggle it on, start a conversation, and point it at your exported ticket file.
4. When you’re ready to run a scored analysis, paste your license key directly into the chat message. Claude uses it only for that conversation and never stores or repeats it back — once per new conversation, not once ever.

Get a license to download the skill package.

Endpoint

https://requirementshub.ai/api/public/evidence/mcp

Speaks MCP Streamable HTTP (JSON-RPC 2.0 over POST). Stateless — clients carry the session token themselves across calls. The same handlers back the REST endpoints under /api/public/evidence/v1/*, so both paths behave identically.

Authentication

Evidence Signal uses bearer license-key / session-token authentication — not the Supabase OAuth flow used by the main SaaS MCP connector. Call evidence_start_session with your license key to mint a short-lived session token, then send that token as Authorization: Bearer <token> and in the token field of every subsequent tool input (the mismatch guard rejects mixed-identity calls).

The session is bound to a seat, and each seat is bound to a RequirementsHub user at provisioning time. That is how an accepted candidate gets a truthful author without a second confirmation step in the app.

What actually gets sent

Ticket bodies are parsed on your machine. What crosses the wire is a skeleton: counts, severities, timestamps, salted ticket-ID hashes and structural links. Two strictness modes are enforced server-side, chosen by the sharing mode stored on the evidence source — a client that claims a different mode is rejected rather than trusted.

• local_only — the strict schema rejects every free-text field. Themes and candidates are scored and ranked, and rendered locally from the templates pack. Nothing readable ever leaves the machine.
• share_accepted — theme and candidate titles are permitted, because materializing a requirement into a project requires a title. This is the only mode in which evidence_accept_candidate can write.

Run telemetry is aggregate-only (counts and finding-code rollups) and is retained for benchmarking; per-call usage events are purged on a 12-month rolling window.

Setup

1. Obtain an Evidence Signal license key — see pricing.
2. Register the endpoint above in your MCP client (Claude custom connector, Copilot agent, ChatGPT app, or any MCP-compatible runtime).
3. Call evidence_start_session with the license key and a machine fingerprint. Store the returned token client-side.
4. Fetch the keywords, clustering, and templates packs, cluster your tickets locally, then submit the skeleton.
5. Review the ranked candidates and accept the ones worth keeping with evidence_accept_candidate.

Available tools

• evidence_start_session

  Exchange an Evidence Signal license key and machine fingerprint for a short-lived session token used as the bearer credential on subsequent Evidence Signal tool calls.

• evidence_get_catalog

  Fetch the Evidence Signal catalog: available packs, finding codes, and engine / pack version metadata.

• evidence_get_pack

  Fetch one versioned, watermarked JIT pack — 'keywords', 'clustering', or 'templates' — used to cluster tickets and draft candidate requirements locally.

• evidence_list_target_projects

  List the RequirementsHub projects the seat's bound user may materialize candidate requirements into. Never auto-creates a project.

• evidence_submit_skeleton

  Submit an anonymized Evidence Signal skeleton (structured signals only, no ticket text in local_only mode) for scoring. Returns clusters, themes, ranked candidate requirements, and findings.

• evidence_accept_candidate

  Materialize a candidate requirement into a RequirementsHub project as a real requirement, optionally overriding type and priority. The only write tool; requires a share_accepted source.