Council of AI GSPC (free) MCP Server

Verify one measurement capsule. Pass capsule_json as the capsule's JSON TEXT (exact: number lexemes are kept) or as an object. Recomputes capsule_id (sha256 of the canonical JSON without capsule_id), picks the rule by the capsule's schema (csoai.measurement-capsule/0.2: RFC 6962 Merkle with 0x00/0x01 domain separation; the superseded v0.1 capsule schema: the v0.1 rule), finds the published batch of the same kind, recomputes that batch's root from its published leaves, and returns the audit path of this capsule against the root named by the signed index. States: INCLUDED, NOT_INCLUDED, ID_MISMATCH, UNCHECKABLE, NOT_PUBLISHED. Verifying is free forever. It proves the bytes were published and bound; what they measured is the capsule's own measurement_state and limitations — measurement, not endorsement.

Verify a signed gspc.measurement-card under the published rule (https://councilof.ai/signed/HOW-TO-VERIFY.md): recompute the id from the canonical body bytes, then check the Ed25519 signature under a key PINNED in this verifier and published in the did:web:csoai.org DID document: #card-attestation-1 (the card key of the signed card index) or #card-attestation-2 (the card key added on rotation, 27 Sep 2026; a card names the key it was signed under), plus the board key for the cards it signed. pinned_key in the result names the key that matched. A card that carries its own key proves only that the file is self-consistent — anyone can alter a body and sign it with a key they just generated — so an inline key outside the pinned set is reported INVALID, and a DID key reference outside it UNCHECKABLE. Three verdicts, never two: VALID, INVALID (with the reason), or UNCHECKABLE when the check could not be completed — 'could not check' is a different claim from 'forged'. Accepts the card as a JSON object, a JSON string, a councilof.ai / csoai.org URL, or a 64-hex card id (the record id shown on the site and the card field of list_cards), which resolves to https://councilof.ai/signed/cards/{id}.json. Never a certification.

Check a sha256 against the live public-root merkle via GET /api/proof?sha=. Three states only: VALID (included), INVALID (not a leaf), UNCHECKABLE (proof endpoint unreachable). Does not claim Ed25519 unless sig_ed25519 is present and checked separately. Never a grade.

One axis row from the live GSPC board at https://councilof.ai/api/gspc — every axis the board carries, behavioural and financial families alike, addressed by the axis id exactly as the board spells it: n, accuracy, interval, MEASURED or UNMEASURED status, family, kind, the bank or run-artifact URL behind the row, and dates. An unmeasured axis is a first-class answer — a declared slot with no run behind it, published so the gap is visible — never an error and never a zero. Never a certification.

Live GSPC board totals from https://councilof.ai/api/gspc. Returns the slot count and the measured count as two labelled numbers WITH their kind — a slot is a declared position on the board, a measurement is a real run behind it; the two are never summed and never swapped — plus the board's separation line, and as_of dates for the board and for this fetch. Compact by default; pass detail "full" for the long-form count grammar, the per-family counts and the board's full measured_on block. We measure, never certify. If the board cannot be fetched the answer is a distinct UNREACHABLE state: no cached number is ever presented as live.

GET one public-root card-v0 leaf by sha256 (64 hex) from https://councilof.ai/cards/{sha16}.json. UNMEASURED cells stay visible. States: VALID (fetched); NOT_IN_THIS_CORPUS (the id is in the signed card index, not the public root, so use verify_card on it); INVALID (in neither the live root nor the signed card index); UNCHECKABLE (a source could not be read). Not a GSPC measurement-card.

The published signed-card index (https://councilof.ai/signed/card_index.json): what the index declares (n_cards) and how many rows it actually carries, reported next to — never reconciled with — the count the card store endpoint (https://councilof.ai/api/cards) reports for itself. Two labelled numbers from two surfaces; if they disagree, this tool shows the disagreement rather than picking one. Optional filters return recent rows: axis, limit. Each row carries card_url, the signed body; pass it, or the row's 64-hex card id, to verify_card.

For ONE obligation (article-50, article-53 GPAI transparency, dora or cra) and an optional subject: the obligation record, its counsel-gate status and the already-signed measurement cards that are relevant to it (count plus the first cards, each with its verify link), read live from https://councilof.ai/api/evidence-bundle. Relevant-to is never a determination: no answer says satisfied or not satisfied, and zero relevant cards is answered EMPTY, never as an error and never as a clean bill. Article 53 output is evidence for review, not a legal determination. Evidence for obligation work — not a conformity assessment, certification, audit opinion or compliance determination; grants no status. Measurement, not certification. Verification of any card is free at https://councilof.ai/gspc-verify.

GET the latest MCP handshake trust snapshot (https://councilof.ai/interop/mcp-trust/latest.json): counts of how many internet-facing MCP servers answer a correct initialize handshake, how many respond with an auth challenge, and how many are unreachable. Counts only by doctrine — host details withheld by design. A partial round or a cap change is disclosed in the snapshot. Measurement, never certification. Three states: VALID (fetched), UNREACHABLE (could not fetch), UNCHECKABLE (body unreadable).

Read the latest signed measurement-capsule index published at https://councilof.ai/measurement-capsules/latest.json: the index root over every capsule, each batch (adapter, kind, capsule count, measurement states, batch Merkle root, record sha256, record signature and OpenTimestamps state), the index's own board signature re-verified here against the pinned did:web:csoai.org#board-attestation-1 key, and the anchor states published beside it (OpenTimestamps, Rekor, XRPL) — PENDING is never called attested. States only: measurement, not endorsement; no verdict, score or ranking. NOT_PUBLISHED when no index is served; UNREACHABLE when the source could not be fetched.

GET the permissionless public-root at https://councilof.ai/root.json. Returns merkle_root, card_count, as_of, and UNMEASURED notes. Separate from GSPC. Three states: VALID (fetched), UNREACHABLE (could not fetch), UNCHECKABLE (body unreadable). Never a certificate.

GSPC Route (free, decide-only). Applies YOUR policy and the fixed GSPC floor to a candidate set and returns the chosen candidate, the basis of the choice and an unsigned route record (csoai.evidence-event/0.1, profile csoai.route-evidence/0.1). Candidates are the GSPC MCP tools by default, or ones you declare: your models, your local GPU, MCP tools, A2A agents, x402 resources. Policy is Cedar: presets read-only, local-only, eu-only, no-unmeasured, plus forbid_providers and allow_kinds; anything the router does not understand grants nothing. For the quality axis you name it reads the live board (https://councilof.ai/api/gspc). With no candidates declared, the GSPC tool whose purpose matches the task is chosen (basis task_match); when no tool's purpose matches, the answer is UNTESTED and nothing is chosen, never a pick by name. For candidates you declare, routing is not ranking: a choice is called a separated leader only when the board separated that comparison; otherwise it is TIE or UNTESTED and your tie_break decides. A candidate the board holds no number for is UNTESTED, never 0. Sponsor, bid and placement fields are ignored. The task text is hashed, never stored. Nothing is executed, called or charged; mode execute answers NOT_ENABLED. Measurement, not certification.

Trust per server, not totals: every published measurement capsule about ONE endpoint URL across all batches — MCP contract-parity dimensions (AUTH, PAYMENT, PROTOCOL, TOOLS, VERSION), A2A card-signature state, self-parity cells for CSOAI's own doors, and any later adapter (e.g. tool drift) — each with its measurement_state, observed_at, correction_pointer, limitations, batch Merkle root and an inclusion pointer for independent verification. Read from a static per-endpoint shard keyed by sha256 of the normalised URL. An endpoint with no capsule answers NOT_MEASURED with an empty list — never an error and never a clean bill. No verdict, score or ranking: measurement, not endorsement.

GET the latest x402 catalog trust snapshot: counts of how many catalogued x402 resources open a correct 402 challenge vs how many are phantom on the wire. Counts only by doctrine — host details withheld; a 402 is NOT delivery; measurement, never certification.