Let Claude Code or a small worker on your own machine pick up Autopilot work. Your files, your tools, your model. Syncflow never connects to your computer.
Why you would want this
A coding step is not text. Doing it properly means reading the repository, running the tests and looking at what broke. None of that is available to a server that only has the words you typed.
A local executor is your own computer doing the step, with your files and your tools, and handing the result back for your approval like any other run.
Your machine asks, we never call it
Syncflow does not open a connection to your computer, and it does not need a port, a tunnel or a firewall change. Your executor asks whether there is anything for it, using an API key you minted, and takes the work if there is.
While it is asking, it is online, and your rules can see that. That is what Claude Code is online means as a signal: a rule can send coding work to your laptop while you are at your desk and quietly do something else when you are not.
Two ways to set one up
Both come from the same npm package, so if you already use one you have the other installed.
Claude Code, while you have it open
claude mcp add --transport stdio \
-e SYNCFLOW_API_KEY=sf_your_key \
-e SYNCFLOW_EXECUTOR=1 \
syncflow -- npx -y @syncflowme/mcp-serverSYNCFLOW_EXECUTOR=1 is the opt-in. Without it you get the Syncflow tools and nothing picks up work.
A small worker, for when you are not in Claude Code
SYNCFLOW_API_KEY=sf_your_key npx -y -p @syncflowme/mcp-server \
syncflow-worker --every 60s --model claudeSwap --model claude for --endpoint http://localhost:11434 to run against a model on your own machine, so the text never leaves it at all.
Watch for the green dot
The Executors tab lists every machine that has checked in, when it last did, and which API key it used. Revoking that key stops that machine and only that machine.
What happens to a step that goes local
Exactly what happens to a cloud one. The result comes back as an artifact, the step moves to awaiting approval, and you approve it. Running on your own machine does not skip the approval, and it does not raise the ceiling.
If nothing is online, the rule falls back the way you configured it: to the cloud, down a rung, or to asking you. Work waiting for a machine that never comes back expires rather than piling up.
What leaves the server
The title, description and context of the step your rule picked, and your instructions. That is a deliberate export to a machine you control. It is not sent to a third party, and it happens because you set up an executor and a rule that uses one.
Coming the other way, anything your machine returns is treated as untrusted content. The server stores it and renders it safely. It never runs it.
The full MCP reference
Setup, every tool the server exposes, the environment variables, example workflows and troubleshooting all live on the Claude Code (MCP) page. This page only covers the Autopilot side, so there is one place to fix when the tools change.
Local executors need a paid plan. See the pricing page.