Route the next turn
Choose a hosted provider, local server, account runtime, or compatible endpoint without leaving the canonical thread.
local-first / model-agnostic / MIT licensed
Your local control plane for coding agents.
milim lets you use your own models and subscriptions, keep one canonical thread, review the diff, and ship.
One desktop app. Read the docs.
tool run
pnpm testTests pass after the targeted fix.Workspace context stays attached.The diff is ready for review.
workspacetest outputdiff ready
model switchedthread retained
src/review.ts
+3 −1@@ -42,4 +42,6 @@ reviewDiff- return publish(change)+ const review = await inspect(change)+ if (!review.approved) return hold(change)+ return publish(review.patch)
inline review readyUndoOpen Git
Continue with another model, then review the resulting diff.
how it works
Inside a milim turn.
Choose the runtime, keep the boundary visible, approve consequential actions, and inspect the result without leaving the thread.
Check the boundary
Keep the workspace, destination, Privacy mode, and Approval mode visible before the turn leaves your machine or changes files.
Approve the action
Inspection can proceed while consequential commands pause with their exact arguments and explicit approve or deny controls.
Inspect the result
See changed files and totals, open the full Git diff, or undo the turn from its checkpoint.
quickstart
download, connect, develop.
Install the app
Grab the Windows portable executable or macOS universal disk image from GitHub Releases.
Connect a model source
Add a provider key, sign in to an account runtime, or point milim at Ollama or LM Studio.
Start a dev thread
Make one safe change in Review, inspect the resulting diff, then continue the same thread with another runtime.
download
Open source.Native desktop.
Yours to inspect. Download the Windows portable executable or macOS universal disk image from the latest GitHub release. Linux packaging is not a primary release artifact yet; the Rust server and Tauri app remain source-buildable; docs cover setup.
source firststable download aliases on every release
faq
questions before install.
Short answers for the local-first and provider-boundary parts people usually check first.
Does anything leave my machine?
Local runtimes stay on loopback. Hosted providers are called only when you choose them. New chats default to Privacy Off; choose Redact or Block before sending when that boundary is required.
Which providers?
OpenAI-compatible endpoints, OpenAI, OpenRouter, Groq, Anthropic, Gemini, Replicate, fal, Ollama, LM Studio, Codex, Claude, OpenCode, and Pi are covered.
Is Linux supported?
Windows and macOS are the release artifacts. Linux packaging is disabled for now, but the Rust server and Tauri app can still be built from source.
Is it really free?
Yes. The repo is MIT licensed. Provider usage depends on the keys, accounts, or local runtimes you connect.
Why does Milim connect to Google Drive?
Only to let you choose, view, and work with specific Drive, Docs, Sheets, and Slides files inside Milim. It requests per-file access rather than access to your whole Drive; details and removal controls are in the privacy policy.
How is Milim different from Claude Code, LM Studio, or Open WebUI?
Those tools are strong at a specific runtime or interface. Milim is the local control plane around them: one canonical thread, model hot-swapping, explicit local boundaries, and diff review without rebuilding the work around one vendor.