Sublime MCP Server

Transforme processos da empresa em sistemas autônomos, confiáveis e observáveis. Você define o processo e o resultado esperado; a Engine executa, monitora e comprova o sucesso conforme seus critérios.

Tools

• List doctrine topics

  The topics of Sublime's modelling doctrine — read these before designing an agent. Each entry carries the topic name to pass to read-doctrine and one line on what it decides. This is platform content: it never contains anything your team wrote.

• Read a doctrine topic

  One topic of the modelling doctrine, whole, as Markdown. Addressed by the topic name that list-doctrine returned.

• List selectable models

  Every model a version may declare, with the reasoning efforts it accepts — the same table the publish gate reads to refuse one it does not offer. What each model costs never crosses this border: choosing one is not a billing decision here.

• List configured providers

  Every model provider your team has configured, and when its key was last verified — never the key itself.

• List agents

  Every agent your team owns: its identifier, name, what it is for, whether it is enabled, the version live in production, and whether the working copy has anything unpublished. Start here: the identifiers it returns are how every other agent operation addresses one.

• Create an agent

  Opens an agent with an empty working copy — no version published yet. It links nothing: link its bases and procedures, then publish-agent.

• Read an agent

  One agent whole: its identity, its working copy behaviour, and the complete inventory of every base and procedure it pins — each pin naming the version pinned and whether a newer one exists to adopt. An empty inventory means the agent pins nothing yet.

• Update an agent

  `name`, `description`, `avatarUrl`, `instructions` and `config` are versioned — production keeps reading the old ones until publish-agent; `config` merges by key, so a call only names the group it changes. `enabled` is operational: it turns the agent off at once, whatever the working copy holds. A call naming nothing is a no-op.

• Delete an agent

  Removes an agent. The address is freed for reuse at once; every version this agent ever published keeps resolving for whoever still reads it by version.

• Link a knowledge base to an agent

  Pins a base into the working copy at its version live right now. Linking a base already linked re-pins it at the version live right now — this is how an agent adopts a base it published a new version of. A base with nothing published cannot be linked, and is refused naming it.

• Unlink a knowledge base from an agent

  Stops this agent from consulting a base. The base itself is untouched and other agents keep it.

• Link a procedure to an agent

  Pins a procedure into the working copy at its version live right now. Linking a procedure already linked re-pins it at the version live right now — this is how an agent adopts a procedure it published a new version of. A procedure with nothing published cannot be linked, and is refused naming it.

• Unlink a procedure from an agent

  Stops this agent from starting a procedure. The procedure itself is untouched and other agents keep it.

• List agent versions

  Every version this agent has published, newest first, each by its ordinal, who published it and when — and which one production serves right now.

• Publish the agent

  Freezes the working copy into a new version and opens a publish attempt: the quality gate runs the agent against its scenarios, and production only moves once the attempt concludes and approves. Publishing with an attempt already in flight supersedes it and opens a new one — the earlier attempt never blocks this one. Read how it went with list-agent-deployments.

• List an agent’s publish attempts

  Every publish attempt this agent has made, newest first — whether it is still running, whether it concluded and what the quality gate decided, or whether it was cancelled or superseded by a later attempt. Answers something different from list-agent-versions: a version and the attempt that produced it are not the same question.

• Cancel a publish attempt in flight

  Stops a publish attempt still running. One already concluded, cancelled or superseded is refused, naming the state it is already in. Returns the state the attempt ends in — never an echo of the identifier you sent.

• Roll an agent back

  Moves production to a version already published, without running the quality gate again — operational recovery. Cancels a publish attempt in flight and leaves the working copy untouched. This changes what production runs right now, so only call it once the user has said, in the conversation, that they want it. Refuses, naming it, when the target is already live.

• Restore an agent version

  Rewrites the working copy — its own fields and every pin — from a published version, discarding whatever in it was never published. Production is untouched. This overwrites unpublished work, so only call it once the user has said, in the conversation, that they want it. Restoring the version already live is how you discard unpublished changes.

• Chat with an agent

  Talks to the agent turn by turn, the way a contact would — through the same door a real contact uses, so the turn it arms is the real thing, on the platform's own queue. Omit `versionId` to talk to the working copy; give it to talk to a published fotografia instead. `conversationId` continues an open test chat; omitted, opens a new one and returns its id for the next call. A conversation keeps the subject it opened with — naming a different `versionId` for one already open is refused, naming the

• Read a test conversation

  The full transcript of one test conversation chat-with-agent opened, oldest first: every message, and every turn by its own identifier, which version it ran against (`working` when none), and how its outcome settled. This is where a turn chat-with-agent answered `pending` for is found once it lands.

• List an agent’s test conversations

  Every test conversation chat-with-agent has opened for this agent, in `PREVIEW` — never a production conversation. Each entry names which version its latest turn ran against (`working` when none), when the conversation opened, how many turns it holds, and the last thing said in it.

• List scenarios

  Every scenario an agent owns, with its declaration and the most recent proof it has, or none if it was never measured. Names whether that proof still answers for the draft you are editing, and what diverged when it does not.

• Create a scenario

  Declares a scenario on an agent — no turn or conversation required, and none is created. Checked against the draft you are editing; a violation is refused naming the field. Returns the identifier every other scenario operation addresses it by.

• Update a scenario

  Rewrites a scenario's declaration under the same checks as creating one. The proof it already has never moves with it — a changed declaration is measured fresh the next round.