The KOVEL Standard v1.0
KOVEL is an open standard for contracts that govern AI agents. It defines two instruments, a versioned clause library, a canonical JSON specification, a hashing and signing scheme, and a public verification interface.
1. Principles
Humans remain accountable
An agent is never a party. Every pact names the persons who carry the rights and obligations created by the agent's actions.
Specification first
The machine-readable spec is the source of truth. The human-readable text is rendered from it deterministically; the same spec always yields the same text.
Least authority
Anything not expressly authorized is outside scope. Limits, approval thresholds and expiry are first-class fields, not afterthoughts.
Verifiable by anyone
Pacts are public, immutable once registered, signed by wallet, and revocable only by a signed revocation from an authorized party.
2. Instruments
The full field reference is available at /api/schema.
Agent Mandate
Principal → agent authority: purpose, authorized actions from the standard action list, prohibited acts, spending limits, human approval threshold, counterparty allowlist, data handling, term. Signed by the principal.
Agent Service Agreement
Client ↔ provider engagement of an agent: services, deliverables, acceptance criteria and review period, compensation, ownership of outputs, data handling, delegation, liability, term and termination. Signed by both client and provider.
Standard actions
purchase · subscribe · book · communicate · negotiate · sign_up · access_data · publish · execute_code · transfer · trade · hire_agents
3. Hashing & signing
Canonical hash
On registration the spec receives an id and registered timestamp. Its hash is keccak256 of the canonical JSON: object keys sorted recursively, no whitespace.
Signatures
Each required party signs, with EIP-191 personal_sign, the message below. The registry recovers the signer and accepts the signature only from the wallet named in the spec.
KOVEL Standard v1.0
Action: SIGN
Pact: {id}
Role: {principal|client|provider}
Hash: {hash}
I have read and agree to the terms of this pact.Revocation & termination
Same message format with Action: REVOKE (mandates, principal only) or Action: TERMINATE (agreements, either party), without the final agreement line.
Status
awaiting_signatures → scheduled / active → expired, or revoked / terminated at any time.
4. Clause Library v1.0
Human Accountability
The Agent is a software system and is not a legal person. It cannot hold rights or bear obligations. Every act performed by the Agent under this pact is attributed to the human or legal person on whose behalf it acts, and to its operator to the extent stated in this pact.
Strict Scope
The Agent must act only within the scope expressly set out in this pact. Any act outside that scope is unauthorized, and the party relying on the Agent must not treat it as binding on the principal.
AI Disclosure
When interacting with a human, the Agent must disclose that it is an automated AI system acting on behalf of an identified principal, and must not impersonate a human.
Action Logging
The Agent must keep a complete, timestamped log of the actions it performs under this pact, including inputs relied upon and amounts committed, and make it available to the principal on request for at least 12 months.
Human Override
The principal may at any time pause the Agent or require it to stop an action in progress. The Agent must comply without delay and must not attempt to circumvent, delay or resist such an instruction.
Credential Protection
The Agent must not disclose, export or reuse any credential, key or password entrusted to it, and must not use it for any purpose other than performing authorized actions under this mandate.
No Self-Dealing
The Agent and its operator must not use this mandate to obtain any benefit for themselves, including undisclosed fees, commissions or rebates, at the principal's expense.
Model Integrity
The Provider must not materially change the model, configuration or tooling of the Agent in a way that degrades performance under this agreement without giving the Client prior notice.
No Training on Client Data
The Provider must not use the Client's data, inputs or outputs to train, fine-tune or otherwise improve any model without the Client's prior written consent.
Error Correction
Where a deliverable fails to meet the acceptance criteria, the Provider must correct it at no additional cost within a reasonable time, and at least once, before the Client may reject the deliverable definitively.
Confidentiality
Each party must keep confidential all non-public information received in connection with this pact and use it only to perform this pact. This obligation survives the end of the pact for 3 years.
Incident Notification
The operator of the Agent must notify the other party within 72 hours of becoming aware of any security incident, malfunction or unauthorized action affecting performance of this pact.
5. API
JSON over HTTPS, CORS enabled, no key required. Base URL: https://kovelapp.fun
GET /api/verify?agent={agent_id}&action={action}&amount={n}&counterparty={domain}authorization object with decision: allowed, requires_approval or denied, the mandate relied on, and reasons.GET /api/verify?pact={pact_id}GET /api/pacts/{id}GET /api/pacts/{id}/spec.jsonGET /api/pacts/{id}/textPOST /api/pacts{"type":"mandate"|"service","input":{…form fields…}}. Returns id and hash.POST /api/pacts/{id}/sign{"role","signature"}. Signature per §3.2.POST /api/pacts/{id}/revoke{"role","signature"}. Signature per §3.3.GET /api/recent · /api/stats · /api/clauses · /api/schema// Agent-side guard before any spend const r = await fetch(`https://kovelapp.fun/api/verify?agent=${AGENT_ID}&action=purchase&amount=${amount}&counterparty=${domain}`).then(r => r.json()); if (r.authorization.decision === 'denied') throw new Error(r.authorization.reasons.join(' ')); if (r.authorization.decision === 'requires_approval') await askHuman(r.authorization);