Quickstart¶

This quickstart assumes the AUMP organization repositories are cloned as siblings:

projects/
  spec/
  conformance/
  aump-py/
  aump-js/
  examples/
  site/

1. Run the Conformance Contract¶

The native runner is Go:

cd conformance
go run ./cmd/aump-conformance validate fixtures

Expected output:

AUMP v0.1 conformance v0.1.0 (spec 0.1.0)
29/29 passed

Python parity runner:

uv sync
uv run aump-conformance validate

After the first package release, the same command is available as a globally installed tool:

uv tool install aump-conformance
aump-conformance validate

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.

After package publishing is configured:

uv tool install aump
aump evaluate-action \
  --mandate mandate.json \
  --action action.json

3. Evaluate an Action in TypeScript¶

cd aump-js
npm install
npm test

The TypeScript tests consume the same conformance fixture corpus.

After package publishing is configured:

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 Marketplace Proof¶

cd examples
uv sync
uv run aump-examples marketplace

The proof runs conformance first, then executes a deterministic buyer/seller marketplace flow that validates A2A metadata, MCP metadata, protected disclosure denial, budget denial, checkout escalation, and canonical evidence events.

After package publishing is configured:

uv tool install aump-examples
aump-examples marketplace

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.