bernstein-mcp
bernstein-mcp
Stateless, read-only MCP endpoint at https://mcp.bernstein.run that verifies bernstein run receipts and hash chains. No account, no key, nothing stored, nothing fetched.
claude mcp add --transport http bernstein https://mcp.bernstein.run/mcpWhat it does
Tool | Result |
| recomputes every chain a run receipt embeds, rebuilds the signed subject, checks the Ed25519 signature; verdict + one line per check, plus the verdict as a signed statement (below) |
| the same verification, narrated: what the run recorded, where it diverges, what the result does and does not prove |
| walks journal rows, lineage entries or audit events on their own (pass the file text for byte-exact rows); names the first broken link |
| TRACE v0.2 conformance checks on one Trust Record, stateless, no account: vendored schema, profile rules, the embedded signature with the key in |
| walks a set of Trust Records from the leaf to the root and classifies the chain ( |
| stateless check of a signed agent manifest (v0.2 COSE envelope): vendored schema, profile, canonicalization, embedded signature (Ed25519, P-256, P-384); optional trust record cross-check via |
| how a bernstein run maps onto a TRACE v0.2 Trust Record, claim by claim; pass a receipt to fill in what its journal answers |
| the compliance presets this release ships |
| the agent adapters bundled with this release |
| version and limits |
Same walk as bernstein verify-receipt, reachable from any MCP client or the
paste form at /verify. A verdict's page
address is the receipt's own digest.
Related MCP server: ShipReceipt MCP Server
Trust model
Proves | Does not prove |
the embedded rows are exactly the rows that were signed | that the embedded key belongs to who you think (trust-on-first-use; pin the operator's published key yourself) |
nothing was edited, reordered, dropped or appended after signing | audit-range HMAC values (keyed by the producing install; their linkage and content hash are still checked) |
Pass the receipt file's contents as a string: an already parsed object
cannot tell 1 from 1.0, and the reference hashes the original spelling.
Signed verdicts
Every verdict comes back as a DSSE
envelope (signed_verdict) signed with this deployment's Ed25519 key: the
payload is a JCS-canonical statement naming the receipt by digest, the
verdict, every check, and an appraisal block in the EAR status vocabulary
(affirming / contraindicated / none). Keep it next to the receipt;
anyone can re-check it offline against the public key at
/.well-known/bernstein-mcp/keys.json
(kid = RFC 7638 thumbprint). The signature covers the DSSE PAE of
application/vnd.bernstein.verdict+json, the same construction the receipt
itself uses. It attests that this verifier reached this verdict for these
bytes at this time — nothing about the receipt's producer.
Session receipts
bernstein-attest turns a Claude Code or Codex CLI session into a signed run receipt that this server verifies.
npx bernstein-attest init # registers hooks for both agents (user scope); --claude-code / --codex / --project to narrow
bernstein-attest link # verify link + one Markdown line for a PR description
bernstein-attest verify .bernstein/receipts/<session>.jsonCodex runs project hooks only after they are trusted once with /hooks in an interactive session; codex exec skips untrusted hooks without printing anything. With --project, Claude Code hooks go to .claude/settings.local.json and Codex hooks to .codex/hooks.json; both name the bundle under your home directory, so keep .codex/hooks.json out of version control.
Recorded per tool call: tool name, hashes of its input and output, success flag, repo-relative path for files the call wrote (hash only for paths outside the project), the program name of a shell command (first token, path stripped, at most 32 characters). Not recorded: prompts, model output, command text, file contents, absolute paths. The receipt is sealed at the end of every turn that recorded something new, and every 2 000 rows; the file lives in .bernstein/receipts/ and can be committed with the change it describes. Because later turns reseal the same file, run npx bernstein-attest link right before committing it so the link you paste names the current bytes.
Limits
Request body | 1 MiB |
Rows per chain | 2 000 (larger receipts: verify locally) |
Rate | 60 requests/min per address on |
Logging | one JSON line per request: route, method, status, JSON-RPC method, tool, verdict, producer family, MCP client name/version, country, colo. Never the address, a header, the body or the receipt |
Correctness
vectors/ holds eight golden vectors generated by the Python reference
(scripts/gen_vectors.py against the pinned bernstein source). The
TypeScript verifier must reproduce each vector's verdict, every check, the
binding bytes and the receipt digest byte for byte (test/vectors.test.ts).
vectors/session/ holds three frozen session receipts written by
bernstein-attest (scripts/gen-session-vectors.mjs).
Development
npm install
npm test # no-fetch guard + vitest
npm run typecheck
npm run dev # wrangler devsrc/ may not call fetch, eval or new Function; CI fails otherwise.
Deploying: see DEPLOY.md. Licence: Apache-2.0.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Verify before your agent acts on data it paid for. Signed verdicts, checkable offline, via x402.
RFC 9421 / Web Bot Auth signature doctor: per-check verdicts, digest, freshness, signed receipt.
Issue & verify signed (ed25519), hash-chained, timestamped provenance receipts for agent actions.
Ed25519 receipts over work artifacts: issue and verify offline, no account, no chain, no network.
Related MCP Servers
- FlicenseNot gradedqualityDmaintenanceVerifies signed SAR v0.1 settlement receipts offline using Ed25519 signatures and bundled key registry.-
- AlicenseNot gradedqualityBmaintenanceProvides tools to create, run, and verify deterministic delivery receipts for completion claims, enabling reviewers to check freshness and correctness of evidence.MIT
- AlicenseNot gradedqualityCmaintenanceEnables offline verification that a supplied DSH workflow adhered to declared duty-separation constraints, producing a redacted, content-addressed JSON verdict.MIT
- AlicenseNot gradedqualityCmaintenanceEnables offline, deterministic verification of credential retirement settlements from explicit receipts, proving that rotations activated new credentials, rejected old ones, and closed without exposing secrets.MIT