Release notes

Where Mini-claude stands, version by version. The project is an openly admitted WIP: the agent loop, tools, modes and permissions are in; sessions, MCP and hooks are not yet.

v0.1: the agent takes shape

The first building block: a full local agent, from chat to action.

Done

Agent & modes

• Agentic loop: the model calls tools, Mini-claude runs them and feeds the result back.
• Four modes, auto, explain, guided, review, with on-the-fly switching via /mode.
• Fallback parser for weak-tool-calling models, startup probe that flags weak tool calling.

Tools

• Files: read_file, write_file, edit_file, grep, glob, tree.
• Commands: bash.
• Web & browser, from your own Chrome: browser_open, browser_click, browser_type, browser_key, browser_scroll, browser_fill, browser_read, browser_screenshot, plus web_search and web_fetch.

Safety & reliability

• Permission with a diff before any destructive action, session-wide “always allow”.
• Git auto-commit + /undo that never touches your own commits.
• Loop guard and a marker on answers produced without reading any file.

Comfort

• Commands: /mode, /model, /workspace, /audit, /compact, /clear, /undo, /quit.
• Project memory: MINICLAUDE.md and other tools’ files (AGENTS.md, CLAUDE.md, GEMINI.md, .cursorrules…) injected into the system prompt.
• keep_alive to keep the model warm, a debug log for diagnosis.

Known limitations

• Tool-calling reliability on small models: some emit calls as text. We recover many, but multi-step tasks stay flaky. See Models.
• Confabulation: a model may answer confidently without having read the file. We flag it rather than hide it.
• Context window: long sessions auto-compact; /compact when needed.
• No undo outside git: /undo requires a git repo.

What’s next

Nothing’s set in stone, but here’s what we’re aiming at.

• Sessions, save and resume a conversation.
• MCP, plug in Model Context Protocol servers to extend the tools.
• Hooks, trigger actions at key points in the agent’s lifecycle.
• First release, prebuilt binaries for macOS / Linux / Windows.
• Model benchmarks, repeatable measurements to guide your choice.