mu · Changelog
Changelog
Every desktop release carries a mu-agent inside it; the main changes are listed here by desktop version. All of them are pre-releases: mu is in early development.
- Settings, sorted by area. The sidebar is grouped into Preferences, Models, Kernel, Details, Capabilities and System, and every page is short. Each feature is one row with its switch, its decision points under it, and the search box at the top of the sidebar finds every setting.
- Skills and MCP servers as mu sees them. Skills are listed by source (mu's, the project's, or inherited from Claude Code, Codex or Cursor); add one from a folder or remove your own. Each MCP server has a switch, and keys passed to a server as arguments are masked.
- A new first-run guide. Model, then permissions (Jev approves by default), then the judge, then a start. Each step says what already works, and the last one offers your Claude Code and Codex conversations when it finds any.
- Quieter conversations. What mu tells the model (the reminder to finish, a goal's continuation, new errors, a rewind) is one line in your language with the exact words one click away; a folded run of tools says what it did, such as "read 3 files · ran 1 command".
- mu-agent 0.1.8. Decision points are active by default; with no key, the free Jev on OpenCode Zen answers; three new decision points (38 in all), one of them screening instructions aimed at the model in web pages and MCP results; classifier models can be judges; MCP servers run on pi's built-in client; built on pi 1.0.2.
- Conversations run on mu directly. A new conversation no longer goes through a middle layer: the app starts mu in a process of its own and talks to it over one typed stream. The conversation list is mu's own session files, so a session started on the command line shows up in the app, and the other way round.
- What the native conversations have. A fork button under each message; rename, pin, mark unread, archive and delete from a row's menu; the goal line and the context ring above the send box; messages queued during a run that you can take back; the source tab with the folder's git changes; a notification when another conversation finishes or waits for you.
- mu-agent 0.1.7. A run you stop ends as stopped, not failed; with only a Together or OpenCode Go key, mu starts on Kimi K3.
- Start with an API key. The first-run guide tells the service from the key's shape (DeepSeek, Kimi, Qwen, GLM, SiliconFlow, OpenRouter, OpenAI, Anthropic, Google Gemini, xAI and more), checks it with one request and picks a model to start with. A wrong key, an empty balance, a refused region or an unreachable service each say what to do.
mu setupon the command line. The same flow, in English or Chinese; a firstmuwith no model asks once whether to set one up.- Windows: installed in C:\Program Files, it starts (#4).
- Behind a proxy. With no
HTTPS_PROXYset, mu in the app, sign-in and the key check follow the system's proxy. - macOS: a conversation no longer puts a bouncing icon in the Dock.
- Choose how Jev is reached. The judges page offers TypeSafe, OpenRouter, the Vercel AI Gateway or a service of your own, each with its own key, and CLM as a third judge.
- Your own model servers. A provider can have an
http://address on your computer or local network, and each model's thinking levels can be edited (#1, #2 by @drunkduckdrown). - Light and dark follow the system, and the menu bar, the permission modes and the commands speak the app's language.
- Permission cards sit under the call they ask about and say what was decided.
- mu-agent 0.1.5 on pi 0.87.1 (Claude Opus 5.5, GPT-6 Sol and Luna, Grok 4.7): it asks before more kinds of risky commands, keeps credentials out of what the judge and the board see, and
/hive <question>starts a hive on your question.
- The app updates itself. It offers a newer version and its size, downloads only after you agree, and installs on restart. Coming from 0.1.1 or 0.1.2, install this one by hand once.
- Windows: answering a permission request no longer breaks the conversation; a missing Git for Windows is explained with its download link.
- mu-agent 0.1.4 starts on a server with 1 GB of memory, using about 50 MB of heap.
- The first preview builds. The desktop app for macOS (signed and notarized), Windows and Linux, each for x64 and arm64; it carries mu inside it, nothing else to install.
- The command line on npm:
npm i -g mu-agent. The judgment kernel on pi, withmu doctor,mu ledgerand the local judge Laya. - Import conversations:
mu importbrings in Claude Code and Codex conversations to continue; settings texts in 12 languages.
If mu is useful to you, star it on GitHub
A star helps more people find it. The code, the discussions and every release live in the repository.