# Runners: where Myco's work runs A runner is a machine you choose to lend a signed-in coding agent to Myco. The Deployment stores sessions and queues work; registered runners claim that work and carry it out. A coding machine can capture sessions without being a runner. Running a Deployment does not enroll its host either. Use a machine that can reach the Deployment and stay available long enough to finish tasks. For a loopback Deployment, use the same machine. For Cloudflare, a runner can be elsewhere. Agent sign-ins stay on the runner; its registration is separate from a member's sign-in and does not enable capture. ## 1. Sign in to a coding agent on the runner Install a supported coding agent, such as Claude Code, Codex or OpenCode, and complete that agent's normal sign-in **on the runner machine**, under the account that will run the service. A login on your laptop does not sign in another machine. Check that the agent can actually run work with your subscription or provider access. In **Settings** → **Models and keys**, configure the model tiers for the agent you use. Codex and OpenCode need a model for each tier that tasks use. The **Sign in with** field offers **Server login** and **Worker login**. For a registered runner, use **Worker login** and sign in on its machine: a server-held login does not sign that runner in, and runner claims do not receive server-held provider keys. In **Settings** → **Myco’s work**, **Preferred agent** and **Then try, in order** choose among the agents the runner offers. A listed model alone does not prove access or available quota. ## 2. Register and start it Install the Myco binary using the [installation steps](quickstart.md#1.-install), then run: ```bash myco runner register
--name ``` Use the full `http://127.0.0.1:8787` address for a local Deployment. A bare hostname uses HTTPS. Registration prints a device approval link and code and waits; it works over SSH without a browser on the runner. Open the link in a browser signed in to the dashboard as an owner or administrator. Check the code, runner name and machine, then approve. Start the enrolled runner's per-user service: ```bash myco runner install --server
myco runner status --server
myco runner doctor --server
``` The service starts at login and uses this account's agent sign-ins. Status shows registration, service, contact and offered agents; doctor diagnoses what prevents it from running. To run in a terminal instead, use `myco runner run --server
` and stop it with Ctrl-C. One runner serves a given Deployment per machine; a second executor waits for the first to stop. ## 3. Check the Runners page Open **Runners** in the dashboard. It shows contact, readiness, current work, offered agents, version, release channel and update results. Open **Details** for agent sign-in reports and model information. If it is not ready, check its reported reason and run doctor on that machine. An owner or administrator can **Pause** a runner to let its current task finish while taking no new work, then **Resume** it. **Remove** revokes its authority immediately; it does not terminate the remote process. Unfinished work can retry on another runner after its assignment expires. To stop the local service while retaining registration, run `myco runner uninstall --server
` on that machine. ## Let Myco work The project switches **Context for sessions**, **Code map** and **Learning** are on by default. A connected repository is also needed for code maps and learning from code. These switches do not turn on the Deployment's schedule: in **Settings** → **Myco’s work**, enable **Work on a schedule** for automatic work. Use **Run a task** under **Myco’s work** for a manual start, and check its result there. By default, **two runs of the same task can run at once across all projects**. Change **Runs of one task at once** under **Settings** → **Myco’s work** → **How much at once**. Work past a concurrency limit queues until capacity is available. Adding runners does not override the limit. Daily task ceilings are separate: reaching one holds off another start until a later scheduling check. ### The first-extraction ramp A large first import is learned gradually. When initial eligible history exceeds one extraction batch, automatic learning starts with a limit of **two runs per day per project**. The ramp ends when that initial history is processed or its active window expires (14 days by default). Later sessions do not extend the initial-history window. This daily ramp is separate from the two-at-once concurrency default. Check **Myco’s work** to see completed runs and why further work is waiting. Sessions can already be visible while learning is still catching up. Small histories and projects that have already processed prompts do not need the ramp. ## Updates between runs Installed runner services check for eligible releases every six to ten hours and update between runs. On **Runners**, choose **Update now** when an update is available; a busy runner finishes its run before applying it. The page reports checks, availability, health verification and the last result. On the runner machine: ```bash myco runner update --check --server
myco runner update --server
``` The updater verifies the release before replacement and checks authenticated contact after restart. A failed handoff restores the previous binary. A failed release can be blocked temporarily; the page shows the reason and expiry. **Clear block and update** or `myco runner update --server
` requests a retry. A foreground runner using another executable refuses service updates. Members update on demand with `myco update`. If the same machine also captures sessions and its runner handoff leaves agent refresh pending, run interactive `myco update` or `myco member provision --refresh` afterward. See [Updating Myco](upgrade.md). ## Retire a legacy worker An older standalone worker service stays until you explicitly retire it. On its machine, run `myco worker status`, let assigned work finish, and use the retirement command it prints. Then register and install the runner. Retiring the legacy worker preserves membership and capture; it does not move its agent sign-in to another machine.