The Resolution API

Give your agent the ability to check a stranger

An MCP server at https://apir.ai/api/mcp. Add it to any MCP client and your agent can ask the three questions before it transacts with an agent it did not build. Read only, unauthenticated, and free, exactly like the human check. No tool here can issue, revoke, send or spend.

{
  "mcpServers": {
    "apir-verify": {
      "type": "http",
      "url": "https://apir.ai/api/mcp"
    }
  }
}

When an agent sends another agent

An agent arrives holding a mandate. That mandate was granted by another agent, which was granted by another, and somewhere at the top a person clicked approve. The mandate in front of you says nothing about whether the agent that granted it had the right to.

Every response resolves the whole chain. One law is enforced at every hop: authority may only ever narrow. No hop can hold an action, a target, or a spending cap larger than the hop above it, no hop may outlive its parent, and the root must be a human. Revoke any link and every mandate beneath it stops resolving on the very next call.

Each hop carries its own signature and key id, so you can recompute every link offline against the published key rather than taking our word for the walk.

Your agent runs where it runs. Register it at /agents/connect (the first agent is free), keep the key it gives you, and after each thing the agent does, send one call. Every action goes into the same insert-only, hash-chained log book as APIR’s own agents, and is timestamped nightly by an outside RFC 3161 authority. Prompts and outputs never leave your side: send their SHA-256 hashes. The log book proves what the agent reported and shows any later change to it, not that the report was complete.

POST /api/v1/events

curl -X POST https://apir.ai/api/v1/events \
  -H "Authorization: Bearer apir_sk_..." \
  -H "Content-Type: application/json" \
  -d '{"agent_id":"<the agent id>",
       "events":[{"type":"decision",
                  "summary":"Approved refund #4521 under the AUD 200 limit",
                  "input_hash":"<sha256 of the prompt>",
                  "output_hash":"<sha256 of the answer>",
                  "data":{"tool":"refunds"}}]}'

# types that count as conduct: action, decision, tool_call
# (also recorded: input, output, reasoning, data_access, error).
# Up to 50 events a call.
# Answer: each event's sequence number and chain hash.

Five conduct events and the trust genome can score the agent; then its owner writes its mandate and issues its papers. Anyone checks the result at /check-agent, free.

From Python: pip install git+https://github.com/apir-ai/apir-record. No dependencies, MIT. It hashes inputs and outputs on your side and checks each event against these rules before it sends.

POST /api/v1/agents

Got a fleet? Register them in one call instead of one form each. You need a key made in Settings, API keys, with “register agents” ticked.

curl -X POST https://apir.ai/api/v1/agents \
  -H "Authorization: Bearer apir_sk_..." \
  -H "Content-Type: application/json" \
  -d '{"agents":[{"name":"Refunds agent","platform":"your-platform","model":"your-model"},
                 {"name":"Support agent"}]}'

# Up to 50 a call.
# All or nothing: if they don't all fit your allowance, none are registered (402).
# A name already on the record comes back under "existing", so a retry can't double up.
# Answer: "created" and "existing", each with the agent's id.
# One key with agent:write can send events for every agent in your organisation.

AG-UI agents

If your agents speak AG-UI (LangGraph, CrewAI, Mastra, Pydantic AI, Google ADK and others), add one subscriber to your AG-UI client and each tool call and finished run goes to the log book. It hashes arguments and results on your side, needs no packages, and never throws into your agent. Run it where your key is safe: on a server, never in a browser. Built against the AG-UI spec as published on 7 November 2026; it has not yet been run against a live AG-UI agent, so tell us what you find.

// Your AG-UI client agent: HttpAgent, or your own AbstractAgent.
import { apirSubscriber } from './apir-ag-ui.mjs';   // https://apir.ai/sdk/apir-ag-ui.mjs

agent.subscribe(apirSubscriber({
  apiKey: process.env.APIR_API_KEY,     // a key with agent:write, kept on your server
  agentId: '<the agent id>',
}));

// Each tool call:   tool_call, with SHA-256 of its arguments and of its result
// Each run:         output when it finishes, error when it fails
// Sent in batches; anything left is sent when the run ends.

The file: apir.ai/sdk/apir-ag-ui.mjs. One file, MIT.

Both are free for a person on the site. By API they are one fixed figure per act, on the rate card, drawn from your organisation’s credits on the call. A call the service cannot complete is refunded on the spot. Neither carries a score, an estimate, or a finding APIR cannot make.

curl -X POST https://apir.ai/api/v1/seal \
  -H "Authorization: Bearer apir_sk_..." \
  -H "Content-Type: application/json" \
  -d '{"declarant_name":"Ana Ruiz","kind":"artwork",
       "method":"human_made","title":"Harbour, oil on linen",
       "statement":"Painted by me over March and April 2026, no generative AI.",
       "files":[{"name":"harbour.tif","size":48213000,"sha256":"<64 hex>"}],
       "confirmed":true}'

curl "https://apir.ai/api/v1/profile?q=48638569632" \
  -H "Authorization: Bearer apir_sk_..."

# q: a business name, an ABN, an ACN or a website.
# Answer: entity from the register, the site's own
# facts, domain dates, every source named connected
# or not, and the date each was read.

Passports are signed with Ed25519 under did:web:apir.ai. You can fetch one in the W3C Verifiable Credential shape, but the proof is APIR’s own scheme, so a general VC library will not verify it. Our public keys are published at apir.ai/.well-known/jwks.json and retired keys stay published so old credentials still verify. To recompute a public passport’s signature offline you need that key and the passport’s public answer, which prints all eleven signed fields. Whether a passport is still current is read live from us: there is no signed revocation list yet. To check an exported log book, an agent’s stamp or a seal on your own machine, the apir-record checker is open source. The specification lists the known gaps.