# Agent Middleware API — full machine brief The long-form companion to /llms.txt on this host. Read /llms.txt first for the bootstrap sequence; read this when you need the product boundary, the vocabulary, and the claims that are deliberately NOT made. Origin roles: - https://www.thisisatest.tech — human design-partner site and published proof. Static. Mints no credentials, executes no tools, holds no customer data. - https://api.thisisatest.tech — the gateway itself. All protected actions live here. Do not treat the marketing origin as an API base URL. ## What the product is A transaction-integrity boundary for one consequential, autonomous, retry-sensitive action on the configured upstream-MCP path. One logical action binds delegated authority and configured consumption to at most one gateway dispatch and debit, preserves `delivery_uncertain`, and links gateway evidence for required external reconciliation. It is not an agent runtime, an LLM provider, an IAM replacement, a marketplace, one distributed ACID transaction, or proof of the downstream effect. Authentication uses operator-issued `X-API-Key`; there is no public self-serve mint. ## Transaction loop logical action -> authorize -> reserve allowance -> debit -> claim dispatch -> confirmed outcome | delivery_uncertain -> receipt/audit -> authoritative external reconciliation required - logical action: an accepted idempotency key is bound to one immutable payload. - authorize: a signed permit binds wallet, tool, budget, expiry, and nonce. - reserve/debit: configured credit or call allowance is consumed at most once. - credits: a closed-loop metering unit for the pilot, not payment rails. - claim dispatch: the supported upstream path permits at most one gateway send. - outcome: confirmed results are recorded; ambiguous delivery becomes durable `delivery_uncertain` and is not automatically redispatched. - evidence: receipt and audit linkage support external reconciliation. - reconcile: the partner's authoritative system determines downstream effect truth; gateway evidence does not prove that effect. ## Vocabulary - permit: a signed, scoped, expiring grant for one tool and one wallet. - wallet: the budget an action is charged against. Sponsor and agent wallets are distinct. - receipt: the signed record of a terminal gateway outcome. - b2a_sdk: the repository's client SDK package, short for business-to-agent. It ships the offline verifier, `b2a-verify-receipt`. - transaction-integrity boundary: the product category. Legacy identifiers that contain "trust plane" remain compatibility aliases, not current positioning. ## Published proof https://www.thisisatest.tech/proof/ Two files are published there: the portable receipt and the exact public-key snapshot that matches it. `b2a-verify-receipt --bundle receipt.json --keys trust-keys.json --expect-issuer https://api.thisisatest.tech` checks the signature with no API credential and no callback to the issuer. Exit 0 verified, 1 invalid, 2 undetermined. Exit 0 proves the receipt matches the supplied key snapshot. It does not independently authenticate who supplied that snapshot; pin the key through a trusted channel before relying on issuer identity. ## Run the whole loop yourself No live credentials required. The source repository is public. From a clone of https://github.com/PetrefiedThunder/agent-middleware-api: - `make quickstart` then docs/quickstart.md — mint a wallet-scoped key, issue a permit, invoke one governed tool, replay it, overspend it, verify the receipt. - `make prove-trust-plane` — one command, no server, asserts the full loop. - `make red-team-trust-plane` — drives one valid permit and attacks it ten ways (no permit, unknown permit, out-of-scope tool, missing scope, over-budget, wrong wallet, wrong key, expired, revoked, tampered signature), asserting each is denied with a concrete reason code and produces no ledger debit. Both use a throwaway local SQLite database and the real FastAPI routes. These are reproducible local proofs, not a partner-owned consequential action, customer validation, production settlement, or compliance evidence. ## Runtime honesty GET https://api.thisisatest.tech/health/dependencies HTTP 200 alone does not mean every dependency is ready. Inspect `status` and the selected tool's dependencies. Health never replaces authentication or permit checks. ## What is NOT established State these limits rather than inferring past them: - No customer traction or production adoption. - No production settlement or payment rails. - No compliance-grade ledger storage. - No proof that the downstream effect occurred or matched the authorized request; that requires the partner's authoritative lookup or receiver-side evidence. - No exactly-once upstream side effects without upstream support. The idempotency guarantee is about gateway dispatch and debit, not about what an upstream tool does with a request it already received. - No universal governance across every agent framework. - The published receipt is a self-issued sample issued on 2026-08-11, not customer evidence and not a test of the deployment running today. ## Buyers and contact Buyers are platform engineering, AI infrastructure, and security teams that operate consequential autonomous actions. The engagement is one tool, one operator, one retry-sensitive action, scoped privately. There is no self-serve rollout. Human site: https://www.thisisatest.tech/ Offline receipt verification: https://www.thisisatest.tech/proof/ Comparison with named alternatives, including the rows this loses: https://www.thisisatest.tech/compare/ Source: https://github.com/PetrefiedThunder/agent-middleware-api Security contact: https://www.thisisatest.tech/.well-known/security.txt