KeyHalve verify-MCP

Free, public, no-account MCP server that lets any AI verify KeyHalve-sealed documents — from any platform on the rail (ValidPay, CheckBooks, …). Seal = the door (a platform's paid MCP). Verify = the room (this one, free forever).

• Endpoint: https://mcp.keyhalve.com/mcp (Streamable HTTP, stateless)
• Tools: keyhalve_verify · keyhalve_status · keyhalve_explain — all read-only, no auth

The blindness rule

This server never receives decryption keys. A verify URL carries the holder's key share in the #key= fragment; parseInput discards any fragment before any other logic runs, and the response says so. Verification here covers everything provable without the key:

Reading the sealed contents still happens only in the holder's browser — exactly like the web verifier. The overall verdict fails closed: any failed check → FAILED — DO NOT TRUST.

Design notes

• Zero runtime dependencies. WebCrypto only; the whole protocol layer is hand-auditable. Same reasoning as the pinned-key rail client in keyhalve-website.
• Stateless. No sessions, no SSE, no KV, no cookies; every POST gets application/json. Request bodies are never logged.
• Tenant-neutral. Platforms come from the same manifest data as the web verifier (TENANT_MANIFEST in src/verifier.ts); onboarding a platform = one data entry.
• Fail closed. Unreachable rail, malformed share, partial dual-sign binding, unknown id prefix — all report NOT verified, never a soft pass.

Develop / deploy

npm ci
npm run typecheck && npm test   # 32 tests
npm run dev                      # wrangler dev

Deploys are manual (deploy.yml via workflow_dispatch, same discipline as rail/console). Needs the CLOUDFLARE_API_TOKEN repo secret; the route mcp.keyhalve.com is a custom domain on the business CF account (same account as the watchdog scheduler).

Directory submissions (Mike-gated)

Submitting to the Claude Connectors Directory / ChatGPT App Directory is an outward-facing step — prepared separately, goes out only on Mike's go.

Listings

Directory-listing assets live in this repo — reuse them, don't invent copy:

• llms-install.md — AI-agent install steps (Cline's AI-driven install; also the canonical per-client snippets).
• glama.json — Glama claim file (maintainers; their live schema is maintainers-only).
• assets/icon-400.png — 400×400 icon (white split-circle glyph on Ink #0E1116, from the brand kit).
• Descriptions must stay byte-consistent with src/tools.ts and pass the approved-claims register (no "split key", no "tamper-proof", no issuer-identity claims).