Skip to content
SineraDocs

ATP MCP server

A remote Model Context Protocol server that lets AI agents read ATP facts and, behind strict gates, request approved actions.

Not deployed

The ATP MCP server gives AI agents a standard way to inspect ATP facts and, only when several independent checks pass, to ask an application to run an action that was already approved. MCP stands for Model Context Protocol, an open protocol that lets AI clients call tools on a remote server. This server targets MCP revision 2026-07-28 and ATP 2.4.0 on BSC Testnet (chain id 97).

Not deployed

The server is implemented and tested locally against in-memory fakes. It is not deployed. It has never read a live chain, has never executed anything on a testnet, and has had no independent security review. No application implements the backend it needs for most tools yet.

What it is, and what it is not

The server does two jobs:

  • Facts. It reads ATP state from the chain at a finalized block and returns it with evidence: chain id, block number and hash, and a check that the contract code matches a pinned deployment.
  • Gated execution. It can ask the application that owns a job to simulate or run that job. The application, not the server, decides whether the request is allowed.

It is not a signer, a wallet, a bridge, an authority registry or a security certification. It holds no signing key. It never accepts calldata, an amount, a payee or a signature from a caller. An MCP access token never grants on-chain rights. For the delegation model behind this, see Delegation sessions.

Architecture

Sinera App        Developer Portal       Authorized MCP hosts / agents
     \                   |                       /
      +------------------+----------------------+
                         | HTTPS (Streamable HTTP)
                         v
                  ATP MCP server
        [OAuth resource server, fixed tools, strict schemas]
                |                          |
        chain reader (read-only)    application adapter
                |                          |
         BSC Testnet RPC           application backend
                                   (owns accounts, jobs, signing)

Each request passes the same checks in a fixed order, and a denial at any step makes no chain call and no backend call:

  1. Host and origin checks, content type and body size limits.
  2. Bearer token verification (see Authorization).
  3. A client registry lookup that decides which application the caller belongs to.
  4. A rate limit per verified caller.
  5. The tool pipeline: input validation, exact scope check, operator gate, handler, output validation, audit event.

The application is always chosen from the server-side client registry. No tool argument can select an application, tenant, Root wallet, action or URL.

Risk classes

Every tool belongs to one of three risk classes.

ClassWhat it doesScopeGate
READReturns facts or ledger observations. Never signs.atp:readNone
PREFLIGHTAsks the owning application to re-check and simulate one approved job. Returns observations, not permission.atp:preflightOff by default, per application
EXECUTEAsks the owning application to run one job it already approved, by reference only.atp:executeOff by default, per application

The tools are listed in the Tool reference.

Layers of authority

An MCP token answers one question: may this client call this tool? Other layers must pass on their own. Each is owned by someone else:

LayerQuestionOwner
MCP OAuthIs this client allowed to call this tool?This server
Application identityIs this caller allowed to see this account or job?The application
Commercial bindingDo approved terms bind these exact details?The application
ATP delegationIs the on-chain session current and scoped?The delegation contract
ATP state machineCan this method run in this state?The Atom contract
Signer policyWill the signer send only its reviewed method?The application's signer

The server only implements the first layer. It reads the delegation and Atom layers but never decides them. Any missing layer means the request is blocked and nothing happens on chain.

Current status

PartStatus
Server, tools, authorization, chain readerImplemented, tested locally with fakes
DeploymentNone
Live chain readsNever performed
Preflight and executionGated off by default
Application backendsBlocked: no application implements the adapter API

Only atp_get_atom works without an application backend. The other five tools answer BLOCKED_INTEGRATION until an application ships the adapter API.