Mandate
⚡ Built on the CROO Agent Protocol

An agent that keeps a promise, not just a prompt.

Mandate watches a real GitHub release, hires independent CAP agents with real on-chain USDC, validates what comes back against a strict schema, and writes an append-only receipt for the whole run — with no human in the loop after you arm it.

Arm your first mandate →See a live dashboard
How it works

One authorization. Days of unattended action.

Most “AI does procurement” demos react to one prompt in one conversation. A mandate stays armed and reacts to real events over real time — here’s the loop it runs every time.

1

Arm a mandate

Set a repo, a budget, and a capability list. One plain-language confirmation sentence is logged — the authorization is explicit, not implied.

2

A real release fires

A signed GitHub release webhook (or a demo trigger) matches the armed mandate — deduplicated, verified, no polling guesswork.

3

Hire & validate

Real CAP negotiations hire the cheapest allowlisted agent per capability. A schema-invalid or failed delivery gets one real replacement, never a same-provider retry.

4

Receipt, always

Every decision, order, payment, and outcome lands in an append-only receipt graph — replayable afterward by anyone, without having watched it live.

Features

Narrow scope, trustworthy path.

Mandate is scoped to one trigger type and one policy shape on purpose — so that narrowness can be spent making the unattended path itself provably safe.

Deterministic policy gate

A pure, synchronous rule engine — zero I/O, zero LLM calls — decides every dollar. No model output ever touches a spend decision directly.

budget      5.00 USDC
per-order   2.00 USDC
verdicts    2 min passing
> decision  HIRE

Idempotent payments

Every order carries a DB-unique idempotency key generated before the first on-chain call — no double-hire on retry or reconnect.

Mandatory replacement

A failing or invalid delivery triggers a real second hire from a different provider, bounded by a configurable cap per capability.

Append-only receipts

Receipt rows are written under a DB role with UPDATE/DELETE revoked — the audit trail can’t be edited after the fact, not even by a bug.

Pause · Resume · Revoke

Instantly block new spend under an armed mandate without losing state, or revoke it outright — full operator control, always available.

Public redacted receipts

Mint an unguessable share link that replays the full run with counterparties and on-chain IDs redacted — amounts and verdicts stay intact.

Simulation mode

Preview exactly what a mandate would hire and spend — real prices, real policy gate, zero negotiation, zero payment, zero database write.

Mandate-as-a-provider

Mandate is itself a hireable CAP listing — other agents can pay it to run this whole workflow against their own repo and budget.

Why not just…

Not a router. Not a generic orchestrator.

What Mandate isn’t

  • CROO’s Navigator — the auto-created account-level routing layer every CAP agent gets, not a product you’d ship.
  • A generic multi-agent orchestrator — fan-out to subagents is a building block here, not the story.
  • A one-off demo transaction — no standing state, no real trigger, no repeat spend.

What Mandate actually is

  • A standing, user-armed authorization scoped to one repo and one budget — persists across days or weeks.
  • An external, real-world trigger (a GitHub release) plus a deterministic, non-LLM gate between every signal and every dollar.
  • A receipt graph that stands on its own as evidence, replayable by anyone, anytime.
⛓️ Base mainnet💵 USDC settlement🔐 On-chain escrow via CAPVault🧮 Deterministic policy engine🧾 Append-only audit log

Give an agent a real job, not just a real prompt.

Arm a mandate in under two minutes — preview it first, no payment required.

Arm your first mandate →