Changelog and roadmap
What version 0.1.0 of the ATP MCP server contains, what is still open, and what unblocks each milestone.
This page lists what the ATP MCP server contains today and what must happen before it can run for real.
Not deployed
Current status
Version 0.1.0 is unreleased and not deployed. It has been tested locally against in-memory fakes only. Execution is gated off by default. No application implements the adapter API yet, and there has been no independent security review yet. "Implemented" on this page means the code exists and passes local tests, not that it ran on a live system.
Version 0.1.0 (unreleased)
The package is private. It has not been published or released.
Added
| Area | What exists |
|---|---|
| Server | A remote MCP server for protocol version 2026-07-28, built on the official TypeScript SDK v2. It uses Streamable HTTP in JSON response mode, rejects the legacy handshake and keeps no session state. It has a /healthz endpoint and shuts down gracefully. |
| Configuration | 19 ATP_MCP_* variables, validated at startup. Invalid values fail closed and are never echoed. Gates default to off. |
| HTTP guards | Host and Origin allow-lists, exact CORS, bounded request bodies, content-type checks, hardening headers and a correlation id. The bearer token is stripped before the SDK sees it. |
| OAuth resource server | JWT validation (issuer, audience equal to the resource URL, fixed JWKS, asymmetric algorithms only), bearer challenges, protected-resource metadata and a server-side client registry that selects the application. |
| Policy | A fixed table of scope, risk class and rate cost per tool, exact-scope matching and operator gates for preflight and execution. |
| Six tools | atp_get_atom, atp_get_delegation_session, atp_get_execution_attempt, atp_simulate_authorized_action, atp_request_authorized_execution, atp_get_execution_result, each with strict input and output schemas. See Tool reference. |
| Chain reader | A read-only reader for the pinned BSC Testnet deployment (chain id 97). It pins the finalized block by hash, checks staleness and block replacement, and verifies runtime code before decoding. |
| Application adapter | A client for the proposed adapter API, plus a blocked backend that answers BLOCKED_INTEGRATION. |
| Telemetry | Allow-listed JSON audit lines, salted subject pseudonyms and free-text redaction. |
| Rate limiting | An in-process token bucket per caller. See Deploy and operate. |
| Schemas and examples | Generated JSON Schemas with integrity hashes in schemas/tools/, an OpenAPI 3.1 proposal for the adapter API, and synthetic examples in examples/. See Schemas and Examples. |
| Deployment plan | A Node 22 container definition and a written plan. No deployment exists. |
Hardened before release
An independent review pass of the code led to these changes before any release:
- One caller-identity rule. The token verifier, the configuration check and the adapter client now share one grammar: visible ASCII without spaces, 1 to 255 characters. A token with a non-conforming subject or client id is denied with HTTP 401 and no downstream call.
- Not sent versus outcome unknown. If an execution request fails before anything leaves the server, it reports its own category (
EXECUTION_NOT_SENTand similar) and is never retry-safe. Every other failure on that path isRECONCILIATION_REQUIRED. See Preflight and execution. - Node entry tested as run. Tests use the same server construction as production. A raw-socket test proves that duplicate
Authorizationheaders are refused. - Safer startup log. The log reports whether an RPC is configured instead of its origin, because providers embed keys in hostnames.
- Honest examples. The cross-tenant example shows the real
REFERENCE_NOT_ACCESSIBLEcode. - Documented limits. The security model now states the limits of number-pinned chain reads, existence-only
ATOM_OBSERVEDand permissive tokentypand multi-audience handling.
Not done yet
- No live chain read. No network available during development could reach a BSC Testnet RPC.
- No testnet execution.
- No run of an official MCP conformance suite and no third-party host test.
- No application implements the adapter API. Sinera has no service route or machine scope for it. Dashboard has no job ledger, signer or service API.
- No deployment, no independent security review and no incident drill.
Roadmap
Milestones are listed in order. Status reflects the plan's latest state.
| Milestone | Status | What unblocks the rest |
|---|---|---|
| Lock sources and dependencies | Done | Re-verify when an upstream revision moves. |
| Transport and strict schemas | Done locally | A recorded official conformance run and a third-party host test. |
| OAuth and isolation | Done locally | Validation against a real authorization server and JWKS endpoint. Tenant isolation proven with a real backend. |
| Canonical read tool | Partly done | A live read of the pinned deployment from a network that reaches the RPC. |
| Reviewed preflight | MCP side done, blocked | Sinera implements and approves the adapter API: a service route, a machine scope and account mapping. No endpoint may be guessed. |
| Bounded Sinera testnet execution | MCP side done, gated off, blocked | Owner authorization, a backend route that hands off to the existing signer, signer release evidence, durable idempotency proof and a controlled testnet end-to-end run. |
| Dashboard adapter | Blocked | Dashboard owners build a service API, tenant model, job ledger and executor, with separate credentials. |
| Release operations | Partly done | An independent security review, deployment attestation, an incident drill, monitoring and a signed operator review. |
Most remaining work sits outside the MCP server's own code: a live read, application-owned adapter implementations, owner authorization, a security review and deployment. Production activation is a separate decision. Until each relevant gate is independently satisfied, write tools stay disabled.
What this means for you
You can read the code, run the server locally and build a client against the tool schemas. You cannot connect to a hosted server, and you cannot execute anything through it. To try it, start with Getting started. For the wider picture, see the MCP overview and compatibility.