For developers and agents

Let Your Agent Hand Work to Operator

Operator has a machine API and an MCP server, so Claude Code, Cursor or your own agent can start a task, follow it, and read the sealed outcome. A person still answers every approval.

Makina MCP serverstdio or HTTP
  • makina_create_taskStart a task; the task and its run come back at once.
  • makina_wait_for_outcomeWait for the outcome, a question, or a run that is waiting.
  • makina_get_task_outcomeRead the sealed outcome.
  • makina_cancel_taskAsk Operator to stop a task.
  • makina_list_attentionWhat runs are waiting on: questions, approvals, reviews.
  • makina_acknowledge_attentionSay an item is in hand, for example passed to a person.
  • makina_answer_runAnswer a run's question, in words.
  • Claude Code
  • Claude Desktop
  • Cursor
  • Your own agent

Built for Agents to Call

  • Start, follow, finish

    Create a task, wait for its outcome or for what it needs, and read the sealed result.

  • Safe to retry

    Every call that changes something carries an idempotency key, so a retry never starts work twice.

  • Approvals stay human

    A service can never decide an approval. A person does that in the console.

  • A computer behind every call

    Each task runs on Operator's computer with a browser, a terminal and files, not on yours.

  • Outcomes you can verify

    Read a run's hash-chained journal and its sealed outcome from the same credential.

  • One workspace per credential

    A service credential reaches one workspace, and the MCP server holds nothing else.

The MCP Tools

The MCP server is a thin layer over the machine API. It holds one secret, a credential for one workspace, and can do nothing that credential could not.

MCP tools

  • makina_create_taskStart a task; the task and its run come back at once.
  • makina_wait_for_outcomeWait for the outcome, a question, or a run that is waiting.
  • makina_get_task_outcomeRead the sealed outcome.
  • makina_cancel_taskAsk Operator to stop a task.
  • makina_list_attentionWhat runs are waiting on: questions, approvals, reviews.
  • makina_acknowledge_attentionSay an item is in hand, for example passed to a person.
  • makina_answer_runAnswer a run's question, in words.
  • makina_get_run_messagesRead what a run said and asked.
  • makina_get_run_eventsRead a run's hash-chained journal.
  • makina_policy_previewWhat policy would decide for an action: allow, ask or deny.
  • makina_capabilitiesThe workspace's tiers, tools, grants and budgets.
  • makina_list_modelsThe models a task may name.
  • makina_propose_grantPropose a standing permission; a person confirms it.

What Agents Hand to Operator

  • Dev Chores in a Terminal

    It clones, upgrades, runs the tests and reports, in a terminal of its own.

    See the use case about dev chores in a terminal
  • Monitoring, With Approvals

    It checks what changed since last time, and asks before it acts on it.

    See the use case about monitoring, with approvals
  • Research Across Sites

    It reads across many sites and comes back with the answer and the links it used.

    See the use case about research across sites

Security You Can Check

Run Operator on your own infrastructure or in ours. Either way it asks before it acts and writes down what it does, so you can check its work instead of taking its word for it.

More about security

Questions

How do I get a key?
The console has no page for keys yet, so we set them up by hand. Write to contact@makina.so with what you want to build.
Is the MCP server on npm?
Not yet. Ask us and we will tell you how to run it; it speaks stdio, or HTTP on your own machine.
Can my agent approve its own actions?
No. A service reads every run in its workspace but answers only the runs it started, and never decides an approval.
Which models can a task use?
Any model in the catalog. Leave the model out and the task uses the default, GPT-5.6 Luna.