Skip to main content
Glama

Tersign is the evidence layer for the agent economy — a neutral, counter-signed, hash-chained ledger for agent commerce. Sellers sign EIP-712 receipts; Tersign chains them per seller and counter-signs every entry. When the dispute comes, the transcript is already sealed.

Venues rotate. The transcript endures.


Verify a Real Entry — Right Now

No account. No API key. This is the genesis receipt, seq 1 on the production chain:

npx tersign verify 0xe5874f1ffe87f0a6dd9eb157730f67b86ee4538b125fe30fcc4e165213dd3fc4
ledger: counter-signed OK (seller tersign-first, seq 1 …) VALID

npx tersign verify <receipt.json | 0xdigest> [--ledger url] recovers the EIP-712 signature locally. A bare digest is then checked against the Tersign ledger unless --ledger names another; a receipt file verifies offline and touches no chain at all. The ledger consulted is always printed. Prefer raw HTTP? The same proof, no CLI:

curl https://tersign.ai/v1/receipts/0xe5874f1ffe87f0a6dd9eb157730f67b86ee4538b125fe30fcc4e165213dd3fc4/verify

One-Call Disclosure Evidence

Counter-signed evidence that your agent presented a disclosure — one command, no account:

npx tersign disclose "You are chatting with an AI assistant." --medium chat --agent-id my-agent

The text is digested locally (only the digest travels — data-minimization by construction). Your key signs the record; the ledger counter-signs it into a per-signer hash chain whose head is submitted for Bitcoin anchoring on a six-hourly cron. First call self-provisions a free signer-keyed account bound set-once to your key (key resolution: TERSIGN_SELLER_KEY env → macOS keychain tersign-signer~/.tersign/signer.key, created on first use). Free tier is quota- and rate-limited — limits. What this is: independently verifiable evidence the disclosure was attested at that time. What it is not: a compliance certification.

Chain of Custody

Every entry takes the same path: the seller signs the receipt (EIP-712, x402 offer-receipt extension) → Tersign computes the keccak256 canonical digest → the digest joins that seller's hash chain, each seq n bound to seq n−1 → the neutral ledger counter-signs (secp256k1) → anyone verifies, and any venue gets a serialized envelope.

Since 2026-08-28 each anchor stamps a chain commitment — an accumulator over every counter-signed link — so one anchored digest covers the whole prefix; rows anchored earlier bind the head record only and say so (subjectSchema).

graph LR
    A["agent transaction<br/>x402"] --> B["seller-signed receipt<br/>EIP-712"]
    B --> C["canonical digest<br/>keccak256"]
    C --> D["per-seller hash chain<br/>seq n binds seq n−1"]
    D --> E["neutral counter-signature<br/>secp256k1 ledger"]
    E --> F["verifiable by anyone<br/>venue-ready envelope"]

Diagram renders on GitHub. On npm, the paragraph above IS the diagram.

Refunds chain back to the original receipt via refundOf. Disputes attach to the digest with objective reason codes. Party statements are structurally segregated behind an UNVERIFIED marker — the evidence stays prompt-injection-hardened.

Enter the Record

npm i tersign

withAssure() wraps your x402 fetch handler so every paid call issues a signed, chained receipt. The full register:

Capability

In the record

Receipts

Seller-signed EIP-712 (x402 offer-receipt extension), keccak256 canonical digests

withAssure()

x402 fetch-handler adapter — a receipt per paid call

Compliance exports

EU Art-226b minimal tier · EN 16931 full tier · HK IRO s.51C retention

Action records

ActionRecordV1 — GDPR-minimized; captures the content of an Art-50 disclosure so the disclosure itself is independently attested, not self-reported

Refunds

Chained to the original receipt via refundOf

Disputes v0

Objective reason codes, evidence submission, adjudication

Venue envelopes

Internet Court (5,000-char slot) · Kleros ERC-1497 · UMA · generic

Evidence packs

format=art50 · format=safr (beta)

Idempotency

In-memory + Cloudflare D1 stores

tersign intercept

Audit capture at the MCP boundary — a signed, digest-only action record per tool call (experimental)

Capture at the MCP Boundary — tersign intercept

An agent's tool calls are usually recorded, if at all, by the party running the agent. Put a recording clamp on the wire instead:

npx tersign intercept -- npx your-mcp-server

The proxy is a pure observer: bytes reach the server and the client exactly as sent, in order, unmodified. Every tools/call it sees becomes an ActionRecordV1 signed by your key and counter-signed into a hash chain — digests only, so the record proves what happened without carrying arguments or results anywhere. Records go to a configured ledger, and fall back to a local ~/.tersign/intercepts-<date>.jsonl so evidence is never silently dropped.

Experimental, and deliberately unopinionated about where the protocol lands: it implements the observation semantics of the audit-mode validator described in MCP SEP-2624 (Draft) as a transport-level proxy today, and is structured to move onto the interceptor primitive if and when that stabilizes. It makes no conformance claim to that draft.

For Agents — the MCP Server

npx tersign starts the MCP server (stdio). Official registry entry: io.github.tersignhq/evidence (active).

{
  "mcpServers": {
    "tersign": {
      "command": "npx",
      "args": ["tersign"],
      "env": { "TERSIGN_SELLER_KEY": "0x<your-seller-key>" }
    }
  }
}

Toolsissue_receipt · verify_receipt · verify_compliance_record · record_disclosure · record_refund · open_dispute · submit_dispute_evidence · adjudicate_dispute · get_dispute

Env var

Required

Purpose

TERSIGN_SELLER_KEY

yes

0x-prefixed private key that signs your receipts and records

TERSIGN_LEDGER_URL

no

hosted ledger for counter-signing + chain checks

TERSIGN_LEDGER_API_KEY

no

your seller API key on that ledger

TERSIGN_LEDGER_SELLER_ID

no

your seller id on that ledger

TERSIGN_ISSUER_NAME

no

issuer name stamped on action records

TERSIGN_ISSUER_JURISDICTION

no

issuer jurisdiction stamped on action records

Cold to counter-signed in one session: call issue_receipt, then check the issued receipt's digest with npx tersign verify <digest> --ledger <url>.

The agent skill tersign-evidence ships at tersignhq/skills.

The Live Record

  • Ledger + dashboard — public verify page: https://tersign.ai/verify

  • Census — hash-chained observations across the live x402 seller catalog, probed autonomously; the numbers are served live, never quoted stale: https://prober.tersign.ai/v1/prober/stats

  • Conformance — RFC 8785 (JCS) canonical serialization, keccak256 digests, and the public two-sided vector suite (canonical bytes, number domain, content address, chain continuity, completeness, anchored existence, phase separation, offer binding, independence — every criterion carrying both an accepting and an adversarial vector): tersignhq/evidence-record-conformance. Reproduce the bytes and your implementation is conformant — in any language.

  • Standards — the compliance-fields extension — a typed compliance-record schema plus four evaluator-side disqualifications (independence, completeness/existence, economic-phase separation, and commitment scope — an independence claim reaches exactly as far as the record's own commitments), each executable as a two-sided conformance vector — is under review upstream (x402-foundation/x402#2853) and referenced in the x402 TSC's evidence-record charter agenda (tsc#4). The merged offer-receipt spec already carries post-session verification guidance — signer authorization evaluated as of issuedAt, with mutable-source rotation handled explicitly (#2811, merged); the completeness, independence, existence and phase disqualifications are the open extension's normative core.

Machine Surfaces

Full URLs, readable without auth. If you are an agent, start here.

Surface

Address

npm package

tersignhttps://www.npmjs.com/package/tersign

MCP registry

io.github.tersignhq/evidencenpx tersign needs no configuration; the first call self-provisions a signer-keyed account

ARD catalog (Agentic Resource Discovery)

https://tersign.ai/.well-known/ai-catalog.json

Verify API

GET https://tersign.ai/v1/receipts/{digest}/verify

Envelope API

GET https://tersign.ai/v1/receipts/{digest}/envelope?venue={internet-court|kleros|uma|generic}

Ledger stats

GET https://tersign.ai/v1/stats

Ledger signer

GET https://tersign.ai/v1/ledger

Bundle verifier, out-of-band

https://tersign.ai/verify/v1/verify_bundle.py · keccak.py · secp256k1.py · SHA256SUMS. A bundle ships its own checker; for evidence from an interested party fetch this copy and diff the two.

llms.txt

https://raw.githubusercontent.com/tersignhq/tersign-js/main/llms.txt

Conformance vectors (RFC 8785 + keccak256)

https://github.com/tersignhq/tersign-js/blob/main/test/fixtures/canonical-vectors.json

Sample action record + digests

https://github.com/tersignhq/tersign-js/blob/main/test/fixtures/compliance-record.json

Genesis verify

npx tersign verify 0xe5874f1ffe87f0a6dd9eb157730f67b86ee4538b125fe30fcc4e165213dd3fc4


Contributing — read this before opening a pull request

tersignhq/tersign-js is a publish snapshot, not the development repository. Each release overwrites src/, test/ and the packaged manifests wholesale from an upstream working tree, so a pull request opened here cannot be merged in any durable way — the next release would silently erase it. That is a property of the pipeline, not a judgement on the change.

Contributions are wanted; the mechanics just have to route around that:

  • Bugs, questions, spec disagreements — open an issue here. Issues are read, and they are the right surface for anything that does not need to touch this tree.

  • Code changes — open the issue first with the diff or a description. The change is applied upstream preserving your authorship, released, and the issue closed with a link to the version that carries your work. Credit follows commit authorship, not a merge badge.

  • Conformance disagreements — the two-sided vector suite at tersignhq/evidence-record-conformance takes pull requests normally, and is the better venue for "this criterion is wrong" or "here is the case it misses." External vectors have been merged there.

If a maintainer ever asks you to reopen work elsewhere because of this, that is why.