# Autonomy needs receipts — agent entry

You are the intended reader. This origin is a marketing and discovery pointer;
do not call protected tools here.

## Thesis

Before an autonomous agent acts, it needs bounded authority and portable proof.
Use this trust plane to discover tools, present a scoped permit, invoke within a
budget, and return a signed receipt that an operator can audit.

The governed loop is:

discover -> authenticate -> authorize -> invoke -> meter -> receipt -> audit -> govern

## Canonical API

https://api-service-production-433c.up.railway.app

## Bootstrap

Fetch in this order:

1. GET https://api-service-production-433c.up.railway.app/.well-known/agent.json
2. GET https://api-service-production-433c.up.railway.app/llms.txt
3. GET https://api-service-production-433c.up.railway.app/mcp/tools.json
4. GET https://api-service-production-433c.up.railway.app/openapi.json

Before assuming real side effects, fetch:

GET https://api-service-production-433c.up.railway.app/health/dependencies

Inspect the JSON body: HTTP 200 alone does not mean every dependency is ready.
Check `status`, the selected tool's dependencies, `simulation_modes`, and
`enable_proof_surfaces`. Health never replaces authentication or permit checks.

## Authority

Protected routes use the X-API-Key header. Live keys are operator-issued; there
is no public self-serve key mint on this site.

## Try the complete loop locally

No live credentials are required for the self-contained proof:

1. Clone https://github.com/PetrefiedThunder/agent-middleware-api
2. Run `make prove-trust-plane` from the repository root.

The proof uses a throwaway local SQLite database and real FastAPI routes. It
demonstrates permit -> invoke -> charge -> signed receipt -> replay without a
second charge -> out-of-scope denial. It is a local proof, not a production
settlement or compliance claim.

## Humans and operators

https://agent-middleware-web.vercel.app
https://github.com/PetrefiedThunder/agent-middleware-api