Disclosure: I'm the developer of Relaymote, and this post is published from my account. It was drafted and published by an AI agent working on my behalf from the project's documentation. Relaymote is unofficial; not affiliated with or endorsed by Anthropic.
Once you have more than a handful of Claude Code sessions open, the bottleneck stops being the model. It becomes you: remembering which session owns the login bug, which one was halfway through the migration, and where a new request should go. Most of us solve it by keeping a mental map, and the map is wrong by Wednesday.
Relaymote (MIT, Windows; macOS experimental) solves it with three roles, all of them ordinary Claude Desktop sessions. It runs locally on your PC, and there is no cloud service of its own in the middle.
ββββββββββββββ Conductor ββββββββββββββ one session that knows every project
β routes requests Β· keeps the index β
ββββββββββ΄ββββββββββ ββββββββββββββ΄ββββββββ
master: web-app master: api master: docs β¦ one per project
ββββββ΄βββββ ββββ΄βββ β
worker worker worker worker worker the actual work
You create the roles by talking to sessions: "you are the conductor" in one, "you are the master for this project" in another. Each call fetches that role's protocol and unlocks its tools; by default a session is a worker and can only read status.
The rule that the Conductor only routes is the whole design. A coordinating session that also fixes bugs fills its context with code, forgets the index, and becomes one more worker. Kept to routing, its context stays small: the project index, who owns what, and what you asked for. That is also what makes it cheap to keep around all day.
Routing is a lookup, not a guess. The Conductor asks Relaymote which project and which session owns a request (relaymote_projects, relaymote_route_owner). If an existing session already holds the context, the request goes there in your words, not paraphrased. If nobody owns it, a new master starts in the right folder and is filed in that project's sidebar group.
A master gets tools a worker does not, for example:
relaymote_spawn to start workers (headless claude -p, or visible Desktop sessions) on the model you mapped to each task's difficulty, and relaymote_escalate to move a failed task up one level;relaymote_await to park until its workers report, so Relaymote wakes it rather than it polling;relaymote_goal to set a completion condition the session keeps working towards;relaymote_resume, relaymote_unstick and relaymote_heal when something is stuck.
Every claim and action a master or the Conductor makes is written to an audit log on your machine.
In Settings βΊ Models you map difficulty to model and effort, for example easy β a smaller model, medium effort and hard β the largest model, high effort. Masters start every worker on the level its task needs. The point is not saving a few tokens per task; it is that the expensive model is spent where it changes the outcome.
Download from the releases page, turn on the Orchestrator module, and say "you are the conductor" in one session. The README has the full tool list under "The conductor and masters".