Mirasim Documentation
Best practices
Ways of working with coding agents in Mirasim: with your team, several at once, on DeepSeek or Kimi, from your phone, and with the cost in view.
Build with coding agents together
Make your computer a collab space (free) and invite people. Start the session you want to steer together on a model everyone can use — a DeepSeek or Kimi key shared with members, or space credit — and turn on “Share with members”. Everyone sends into the same session, and each message carries its sender's name. Personal Claude and ChatGPT subscriptions are for their owner only.
Run several agents without them stepping on each other
Give each parallel session its own Git worktree: each agent works on its own branch, and you review the diffs and decide how to merge. To add to a session that is still running, queue the next prompt or steer the running turn instead of starting over.
Article · Two agents, one repository →
Chapter 6 · Branches, reruns and worktrees › Worktrees: parallel agents on one repository →
Pick the model per session: DeepSeek, Kimi or your own keys
Model and reasoning level are set per session, so two panes can run different models side by side and you can compare their answers. Run agents like Kimi Code and DeepSeek Harness with your own keys, or pick DeepSeek, Kimi and GLM models with Mirasim credit.
Keep going from your phone
Make the computer a collab space, sign in to the phone app with the same account, and pick the computer. Leave “Keep the computer awake” on Always (the default) so the phone can always reach it.
Chapter 19 · Phone, other computers and SSH servers › Use your computer's agents from your phone →
Know which agent needs you without switching windows
Turn on the turn island. It shows your turns at the top of the screen — running, waiting for you or finished — and permission requests and simple questions can be answered right on it.
Know what agentic coding costs
Every model call goes into a ledger with its tokens, cache hits and estimated cost. Look at it per session before a long run, and at the cache hit rate when a session costs more than you expected.
Something wrong or missing on this page? Tell us