A fast, model-adaptive, terminal-native AI coding agent.
You bring your own key. It runs on your machine. Nothing goes anywhere else.
Apache 2.0 • Changelog • Known issues • Contributing • X
Mindweave is a coding agent that lives in your terminal and works inside your repository: reading, searching, editing, running commands, and checking its own work.
It runs entirely on your machine. There is no backend, no telemetry and no account. Your code and your API key never reach a Mindweave server, because there isn't one.
It is built lean on purpose. Most of a coding agent's context budget goes on scaffolding the model never needed. Mindweave keeps prompts thin and leaves the room for the model to reason about your code.
This is Mindweave 1 which is finally out, tagged mindweave-1. Everything in it is here and works from a
clone; the npm package and the website follow shortly. Until then, install from source
with the steps below.
What landed in it is in the changelog: thirteen model providers, a rebuilt terminal interface, reworked prompt caching and token accounting, project notes the agent maintains across sessions, and a long list of things that were quietly wrong.
Requires Windows and Node.js 20+. macOS and Linux are not supported yet, and the reason is written down in KNOWN-ISSUES.md rather than glossed over.
macOS and Linux support is coming soon.
git clone https://github.com/mindweave-cli/Mindweave
cd Mindweave
npm install
npm run build
npm link # makes the `mindweave` command available globallyThen, in any project:
cd your-project
mindweaveIt asks for an API key on first launch and saves it to ~/.mindweave/.env, so every
project afterwards just works. mindweave --help covers the launch flags; everything
else is configured inside a session.
The
mindweavepackage on npm is a name placeholder, not a release. Install from source until the first published version lands.
13 providers, 47 models, one key. DeepSeek, Anthropic, OpenAI, Gemini, xAI, Mistral,
Groq, Cerebras, Qwen, Kimi, GLM, Meta and MiniMax. Each family gets its own driver so it
runs at its best without bloating the shared core, and only the driver you are using is
ever loaded. Switch with /provider and /model; the choice is remembered per project.
Full list: PROVIDERS.md.
Deterministic code intelligence. A background lane indexes your repo with tree-sitter and language servers, costing no tokens, so the agent understands your codebase rather than just the file you opened.
Real tools. File read and edit, multi-file edits, ripgrep search, a shell with
background jobs, sub-agents, and compiler diagnostics. Read-before-edit is enforced, and
/undo is a real net rather than a hope.
Session memory. Long sessions stay sharp through automatic compaction plus a running state summary that survives it. The agent can read its own earlier sessions in a project, so "what did we do last time" gets a real answer instead of a guess.
MCP servers. Connect external tool servers (GitHub, Postgres, your own) with
/mcp add or by asking in plain words. They start with your session, and their output is
treated as untrusted by default. Full guide: docs/MCP.md.
Images and web search. Drag a screenshot into the prompt or write @shot.png; a model
that can see gets the image, and one that cannot says so plainly. Ask about a recent
release and the agent looks it up instead of guessing, through your own provider, with no
second account to manage.
Seeing your app. It can capture one window and look at it, which is how you tell an app that started from an app that works. One window, never the whole screen, and it asks first, naming the window it is about to capture.
Project notes that carry across sessions. MINDWEAVE.md is loaded every session and
maintained by the agent, so a new conversation continues rather than starts over. Split
it with @./path imports, keep machine-wide notes in ~/.mindweave/MINDWEAVE.md, and
put a MINDWEAVE.md inside a folder for conventions that are true only there — that one
is loaded only while the agent works in it. /init writes the first one.
Per-project governor. Give a project standing rules, reusable skills, and forbidden paths or commands the agent has to respect.
| Command | What it does |
|---|---|
/help |
Lists every command |
/init |
Writes MINDWEAVE.md, the project notes loaded every session |
/provider · /model · /think |
Who answers, which model, how hard it reasons |
/clear · /continue |
Start fresh, or resume an earlier session |
/undo |
Reverts what the last turn changed |
/mcp |
Manages connected MCP servers |
shift-tab |
Cycles interaction modes |
Type while it works and your message queues; press up to take it back and edit it.
| CHANGELOG.md | What changed, and what each fix actually was |
| KNOWN-ISSUES.md | What is broken or unfinished, written down rather than carried quietly |
| PHILOSOPHY.md | How the project is run and what gets into the core. Short, and the honest version |
| BOUNDARY.md | What belongs in the core versus a driver |
| CONTRIBUTING.md | Mechanics, and what is most useful to report |
| SECURITY.md | Reporting a vulnerability |
| docs/MCP.md | Connecting and trusting MCP servers |
Contributions are welcome, especially model drivers: Ollama is a stub waiting for someone, and a driver is a smaller job than people expect, owning one provider's wire format and nothing about how the agent behaves.
Small fixes and reproduced bugs with a failing test can go straight to a pull request. For anything larger, start a Discussion first. AI-assisted contributions are fine; the one rule is that you understand and have tested what you are submitting.
Bug reports are genuinely the most useful thing you can send. Mindweave is developed by running it on real projects and fixing what breaks, and nearly everything in the changelog started as a failure someone watched happen. What to include, and what is especially worth reporting.