Mindweave

A coding agent for your terminal.
You bring your own key. It runs on your machine. Nothing goes anywhere else.

Apache 2.0  •  Website  •  Desktop app  •  X

---

Mindweave works inside your repository: it reads, searches, edits, runs commands and checks its own work. There is no backend and no account. Your code and your key only ever go to the model provider you picked.

Prompts are kept thin on purpose, so the context goes on your code rather than on scaffolding.

Names. Mindweave is the program and the name of this repository and of the npm package (mindweave, or mw for short). mwcode is the desktop app: a window around the same engine, for Windows, macOS and Linux, with its own repository and its own download files. The two share sessions and settings.

Install

Needs Node.js 24+. Every change is tested on Windows, macOS (Apple silicon) and Linux. Driving a native desktop app and taking window screenshots are Windows-only for now, and the rest of the limits are in KNOWN-ISSUES.md.

npm install -g mindweave
cd your-project
mindweave          # or mw

It asks for an API key on first launch and keeps it in ~/.mindweave/.env.

From source:

git clone https://github.com/mindweave-cli/mindweave
cd mindweave
npm install
npm run build
npm link

What it does

• 19 providers, 68 models — DeepSeek, Anthropic, OpenAI, Gemini, xAI, Mistral, Groq, Cerebras, Qwen, Kimi, GLM, Meta, MiniMax, Tencent, OpenRouter, local models through Ollama, LM Studio and llama.cpp with no key at all, and a Custom endpoint for any other server that speaks the OpenAI wire. PROVIDERS.md
• Real tools — file reads and edits, search, a shell with background jobs, sub-agents, web search, and a tool that tests your app by using it: it opens a web page or dev server in a browser, clicks through it and looks at the result. On Windows it can also drive a desktop window and take a screenshot of that one window, never the whole screen. It asks first unless you chose auto-accept.
• Code intelligence — tree-sitter and language servers index the repo in the background, at no token cost, and every edit comes back with that file's errors.
• Goals that run on their own — /marathon takes a goal and keeps going until it is verified done, blocked, or out of budget.
• Memory — project notes in MINDWEAVE.md, compaction that keeps what matters, and earlier sessions it can look back at.
• Safety nets — /undo, /rewind, approval modes, per-project rules for paths and commands it must not touch, and a project's own servers and hooks never run unasked.
• MCP servers — connect external tools with /mcp. docs/MCP.md

/help lists every command. shift+tab switches modes. Type while it works and your message queues.

Run it from a script

mindweave -p "task" does one task with no screen and prints the answer, so it works in a shell script or a CI job. Anything piped in is added as context, and --output-format json gives one result object with the session, the tokens and the cost. The exit code says whether it finished (0), failed (1), had no key (3) or stopped before it was done (4). Nobody can answer a question in this mode, so anything that would ask is refused. docs/HEADLESS.md

git diff | mindweave -p "review this change"
mindweave -p "list the TODOs" --output-format json | jq -r .result

Commits

When the agent makes a commit it can add one line to the message, Co-authored-by: mwcode <...>, so the commit shows the agent beside you on GitHub. It asks once per project the first time, remembers the answer, and never adds the line to a commit you make yourself. /coauthor shows or changes the answer, and /coauthor off all turns it off everywhere.

Your own rules and hooks

Both live in your state folder (~/.mindweave, or projects/<name> inside it for one project), never in the repository, so a project you clone cannot add them.

command-rules.md sets what Sentinel asks about, one line per rule, matched on the start of the command:

allow npm test
prompt git push
forbid rm -rf :: use the trash folder

forbid refuses the command in every mode and tells the agent your reason. prompt always asks. allow stops the question for that command, but never for one that writes a file or that Mindweave cannot read with certainty.

hooks.json runs your own commands at set moments. Each gets one JSON object on its input, and exit code 2 stops the action and tells the agent why:

{ "hooks": {
  "PreToolUse": [ { "matcher": "run_command|edit", "command": "node check.js" } ],
  "Stop": [ { "command": "npm test --silent" } ]
} }

The moments are PreToolUse, PostToolUse, UserPromptSubmit, Stop and SessionStart. A hook is stopped after 60 seconds unless it sets timeoutMs.

More

CHANGELOG · Known issues · Philosophy · Contributing · Security

Contributions are welcome, model drivers especially. Bug reports are the most useful thing you can send (what to include). Questions: zallinimann@gmail.com

License

Apache License 2.0