Quickstart¶
This quickstart assumes the AUMP organization repositories are cloned as siblings:
1. Run the Conformance Contract¶
The native runner is Go:
Expected output:
Python parity runner:
The same command is available as a globally installed tool:
2. Evaluate an Action in Python¶
cd aump-py
uv sync
uv run aump evaluate-action \
--mandate ../conformance/fixtures/mandates/marketplace-buyer.valid.json \
--action ../conformance/fixtures/actions/accept-ping-pong.allowed.json
The decision should be allowed.
Installed package:
Compute the portable mandate reference hash:
3. Evaluate an Action in TypeScript¶
The TypeScript tests consume the same conformance fixture corpus.
Installed package:
npx @agentic-user-mandate-protocol/aump validate mandate mandate.json
npx @agentic-user-mandate-protocol/aump evaluate-action \
--mandate mandate.json \
--action action.json
4. Run the Reference Proofs¶
The marketplace proof runs conformance first, then executes a deterministic buyer/seller flow that validates A2A metadata, MCP metadata, protected disclosure denial, budget denial, checkout escalation, and canonical evidence events.
Run the home-buying proof:
The home-buying proof runs conformance first, then evaluates a property search, offer draft, over-budget denial, missing-contingency denial, protected-factor denial, and submit-offer escalation through the same AUMP runtime.
Installed package:
What This Proves¶
Conformance proves implementation agreement on protocol behavior. The examples prove that a deterministic agent runtime can wire AUMP into a real action loop, including outbound A2A mandate references and inbound hash validation.
It does not prove that every LLM will negotiate well. That is a separate application-quality problem. AUMP proves the control boundary around the agent.