Skip to content
SineraDocs

Deploy and operate

The planned container, configuration, health check, logs, rate limits, gates and pre-launch checklist for the ATP MCP server.

This page describes how the ATP MCP server is meant to be deployed and operated. It is a plan, not a record of a running service.

Not deployed

Current status

No deployment exists. There is no hostname, no registered OAuth endpoint and no running instance. The server has been tested locally against in-memory fakes only. Execution and preflight are gated off by default. No application implements the adapter API yet, and there has been no independent security review yet. A running container would not, on its own, make execution ready.

The planned container

The recorded hosting form is a Node 22 container that runs behind a reverse proxy or load balancer that ends TLS.

client --HTTPS--> reverse proxy --HTTP, private network--> container :8080
                                                             |-- egress --> JWKS host
                                                             |-- egress --> chain RPC host
                                                             '-- egress --> backend host(s)
ItemPlan
ImageMulti-stage build on Node 22. The runtime stage holds production dependencies only.
UserThe unprivileged node user. The server needs no writable path.
Listen address0.0.0.0:8080 inside the container, plain HTTP.
Base imagePin it by digest before release.
Other runtimesThe app is a web-standard fetch(Request) => Response. Other runtimes, such as a Worker, are not provided and not verified.

The proxy must forward the original public Host header, leave Authorization and Origin untouched, and apply its own size and time limits. It should expose only POST on the resource path and GET /.well-known/oauth-protected-resource. Only the orchestrator should reach /healthz. The container must not be reachable from outside the proxy.

Configuration

All variables use the ATP_MCP_ prefix. The server validates them at startup. A bad value stops the process with an error that names the variable, never its value.

GroupVariablesNotes
Protected resourceRESOURCE_URLRequired. The public https URL of the endpoint. It is the token audience.
Token validationAUTH_ISSUER, AUTH_JWKS_URL, AUTH_ALGORITHMSIssuer and JWKS URL are required. Algorithms default to ES256,RS256,EdDSA. Asymmetric only.
ClientsCLIENTSJSON map from client id to an application (sinera or dashboard). Empty means every token is denied.
HTTP guardsALLOWED_HOSTS, ALLOWED_ORIGINS, MAX_BODY_BYTESOrigins default to none. Body size defaults to 65536 and is capped at 1048576.
ChainRPC_URLRequired. Https, no credentials in the URL userinfo.
BackendsSINERA_BACKEND_URL with SINERA_BACKEND_TOKEN, DASHBOARD_BACKEND_URL with DASHBOARD_BACKEND_TOKENSet as a pair. If unset, that application's tools answer BLOCKED_INTEGRATION.
GatesPREFLIGHT_ENABLED_APPS, EXECUTION_ENABLED_APPSDefault empty, which means off. Enabling an app that has no backend is a startup error.
Rate limitRATE_LIMIT_PER_MINUTEDefault 60.
ListenerHOST, PORTThe container sets 0.0.0.0 and 8080.
LoggingLOG_SUBJECT_SALTRequired. Salt for the pseudonymised subject in audit lines.

Inject the RPC URL (it can embed a provider key), both backend tokens and the log salt from a secret manager. Never bake them into the image or commit them. The server holds no signer key, wallet key or lease token, and it never forwards the inbound bearer token to a backend. See Getting started for a local run and Authorization for token rules.

Network rules

Ingress is HTTPS only. The app adds its own limits as a second line: your MAX_BODY_BYTES, a 10 second header timeout and a 30 second request timeout.

Allow egress to exactly these hosts and deny everything else, including cloud metadata addresses and private ranges:

  1. The host of AUTH_JWKS_URL.
  2. The host of RPC_URL.
  3. The host of each configured backend URL.

No tool argument can name a URL. The code requests fixed paths on the configured origins and does not follow redirects. It does not itself reject private IP ranges, so the network allow-list is the control. See the security model.

Health, logs and shutdown

GET /healthz returns a constant body and no version or configuration:

{"status":"ok"}

The image health check calls it every 30 seconds.

Logs are JSON lines on stdout. At start you get one redacted startup_config line and a listening line. After that, each audit line carries only allow-listed fields.

Audit fields include the event type, correlation id (the X-Correlation-Id returned to the client), tool, application, client id, a salted sha256: pseudonym for the subject, the decision, closed-vocabulary reason codes and duration.

Logs never contain bearer tokens, prompts, raw tool arguments, service credentials or signatures. To follow one request, search the audit lines for its correlationId. On SIGTERM the server stops accepting connections, drains in-flight requests and exits, with a forced exit after 15 seconds.

Rate limiting

A token bucket applies to each verified caller. Capacity and refill per minute both equal RATE_LIMIT_PER_MINUTE. Every authenticated request costs 1, and a tool call raises its total to the tool's cost.

ToolTotal cost
atp_get_atom, atp_get_delegation_session, atp_get_execution_attempt, atp_get_execution_result1
atp_simulate_authorized_action3
atp_request_authorized_execution5

A refusal is HTTP 429 with Retry-After: 60, or a RATE_LIMITED tool result. Both are safe to retry. The limiter is per process and per replica, and it resets on restart. Enforce quotas at the proxy as well.

Gates and kill switches

Preflight and execution are off unless the operator lists an application in the matching *_ENABLED_APPS variable. Turn a capability off by removing the entry and restarting.

Removing a client from CLIENTS denies all tools to that client. Unsetting a backend URL and token makes that application answer BLOCKED_INTEGRATION. Removing an app from the execution list does not cancel requests already sent to a backend.

These are the MCP server's levers only. The application's own kill switches belong to its operators. Release status today:

GateStatus
DocumentationDone
ProtocolPartial
ReadPartial
PreflightBlocked
Testnet writeBlocked
Two platformsBlocked
ProductionBlocked

Pre-launch checklist

Missing evidence means no-go. Before you turn any gate on:

  1. The owning application implements and approves the adapter API, and issues a credential for this deployment.
  2. A live read of the pinned deployment succeeds from the deployment network, and the evidence matches the pin.
  3. CLIENTS lists only reviewed clients, and your authorization server issues tokens whose audience equals the resource URL.
  4. The egress allow-list matches the rules above.
  5. For execution: the owner has authorized the application and the testnet in writing, the backend's kill switch is tested, and the reconciliation procedure is rehearsed.
  6. The base image is pinned by digest, and secrets come from a secret manager.

See Preflight and execution for what happens once gates are on, and the changelog and roadmap for what is still open.