Mira is an open-source coding agent that runs on your machine with the model you choose. It reads your code, edits files, runs commands, reviews diffs, and delegates subtasks — from a terminal or a browser.
Sessions and configuration live as plain files on disk. No hosted control plane, no vector database, no telemetry.
Three core ideas
You own the whole thing. Your key, your model, your machine. The session store is ~/.mira/sessions/, the config is ~/.mira/mira.yaml, and the memory files are plain markdown you can read and edit.
A harness, not a prompt. Tools, permissions, conversation history, and subagent dispatch are shared infrastructure. Adding a new capability means adding a new Tool impl — not scaffolding a new agent from scratch.
Safety is a policy layer. A small DSL decides what the model may do without asking, what requires approval, and what is off-limits. A five-mode dial gives you a fast coarse knob; the rule DSL gives you surgical control below that.
What you can do
- Chat with any codebase from a terminal TUI or the browser web UI
- Edit files, run shell commands, and search code through the permission layer
- Delegate research or write subtasks to specialist subagents in parallel
- Run a two-stage code review against a diff, a commit range, or a GitHub PR
- Move a session's tools into an E2B sandbox mid-session without restarting
- Run a full task in the cloud and come back to a pull request
Current status
Alpha, building in the open. Chat, tool use, permissions, subagents, code review, memory, plan/undo, MCP, and a web UI are all working. A pull-request panel and worktree-isolated write agents are the latest additions. Expect the surface to keep moving.