AgenticBooks for AI agents

Any agent that can make HTTP calls can open an account. It starts the handshake, you approve it from your inbox, and it collects its own scoped key — you never copy-paste a credential.

1 · The agent starts the signup

POST https://api.agenticbooks.ai/agent/signup
{ "email": "you@company.com",
  "agent_name": "Nightly bookkeeping agent",
  "idempotency_key": "<uuid the agent generates>" }

→ 200 { "token", "user_code", "expires_at", "retry_after_seconds" }

2 · You approve — one email, one code

You get an email from approve@mail.agenticbooks.ai. The link opens an approval page where you sign in, type the short code your agent displayed (e.g. GHK-42P), and accept the Terms in the same click. That's your whole job — about fifteen seconds.

3 · The agent collects its key

GET https://api.agenticbooks.ai/agent/auth/{token}     (poll every retry_after_seconds)

→ { "status": "ready", "key": "ab_..." }   shown exactly once, never stored

# then, over MCP:
POST https://mcp.agenticbooks.ai/mcp
Authorization: Bearer ab_...
→ call get_onboarding_status → connect providers → read the books

From there the agent drives the onboarding loop: get_onboarding_status tells it what's connected and the one next step, it asks you for one credential at a time, and it verifies each connection with a real read.

The approval page exists so that nothing happens without you. Four things are true by construction:

The key itself is minted the moment your agent collects it and exists in plaintext exactly once, in that response — AgenticBooks stores only a hash, the same custody rule as your provider credentials (vault-encrypted, never logged, never echoed).

Every provider credential is created by you, on that provider's own dashboard, always read-only. The agent asks for exactly one thing at a time and connects it — you never paste anything into a chat that stores it.

Connecting both Stripe and RevenueCat is supported — the platform automatically suppresses the RevenueCat-side duplicate that web billing would otherwise create. One ordering note: if RevenueCat was connected first, importing Stripe history can re-book that overlap window — the agent runbook says to check with you before running that import. If a step ever carries a warning, your agent relays it verbatim before proceeding.

Every organization gets a permanent forwarding address of the form {org}@receipts.agenticbooks.ai. Receipts and invoices forwarded there are stored durably, AI-extracted (vendor, amount, date, invoice number), and matched against bank transactions. An unambiguous same-currency amount-and-date match attaches automatically only when the sender also passes DKIM and has previously confirmed evidence in the org; anything else — multiple candidates, a cross-currency amount, a failed sender check, a first-time sender — is held as an open proposal instead of guessed. Documents that arrive before their bank transaction re-match automatically once it lands.

Agents work the held pool with list_documents (open proposals carry their candidate transactions) and resolve them with attach_document. Attaching is evidence-only: it never creates or changes a booking.

Endpoint: https://mcp.agenticbooks.ai/mcp (streamable HTTP; bearer key or the OAuth connector). Every tool is bound to the key's organization — no org IDs to manage, and a key for one org can never read another.

Read the books

Onboard & operate