A coding agent that lives in your terminal.
Bring your own key. Fourteen providers, fifty-three models, switchable per project. It reads your repository, runs your commands, and shows you every change before it lands.
$npm install -g mindweave
v2.3.0
Windows and Node 20+. macOS and Linux are not supported yet.
● Read 1 file └ Read src/drivers/providerError.ts (182 lines) ● A spent balance and a rate limit arrive looking the same. The classifier recognises that the account was refused and quotes the provider's own sentence rather than guessing which one it was. ● Update(providerError.ts) ✓ 0 · 118ms ⎿ - const ACCESS = new Set([401, 403]); + const ACCESS = new Set([401, 402, 403, 429]); ● Run(npm test) ✓ 2,271 passed · 41.9s ● Worked for 34s · 8.2K tokens > say something… tip: /undo restores the last checkpoint
What it does
every item below ships todayFourteen providers
DeepSeek, Anthropic, OpenAI, Gemini, Qwen, Kimi, GLM, xAI, Mistral, Meta, MiniMax, Groq, Cerebras and Tencent. /provider switches; the choice is remembered per project.
Reads the repository
A background lane indexes your code with tree-sitter and language servers, costing no tokens, so it opens the right file instead of grepping around for it.
Every change is a diff
Edits arrive as a diff you read in place, with the language server's own errors for that file attached. Checkpoints are taken as it works, and /undo rolls the last turn's file changes back.
Runs your commands
Tests, servers, builds. A long command moves to the background rather than timing out, and the agent picks the results back up. A recognised test run shows its counts, not a wall of passing lines.
Prompt caching, everywhere
The request is split so the stable half caches on every provider that supports it. A multi-step turn re-sends its context at a fraction of the price.
Rules it has to follow
Per-project rules, skills, protected paths and forbidden commands. Ask for one in plain words and it writes the rule itself.
MCP servers
/mcp add connects a server and its tools join the session. Resources and prompts come with it.
Plan before it builds
In Architect mode it researches and presents a plan, and changes nothing until you approve. Approving starts the work in the same turn.
Updates in place
/update installs the newest release and reopens on the exact conversation you were having, rather than leaving you to close the terminal and start over.
Start in four steps
from npm, or built from sourceInstall it
npm install -g mindweave. Needs Node 20 or newer. Building from source instead is below.
Run it in your project
mindweave, or mw for short. It roots at the current directory.
Paste a key when it asks
On the first run it asks for one and saves it. Use /provider to switch; if it does not have that key yet, it asks for that one too.
Say what you want
Describe the change. Watch the tool calls, read the diffs, and press shift+tab to change how much rein it has.
$ git clone https://github.com/mindweave-cli/Mindweave $ cd Mindweave $ npm install $ npm run build $ npm link
Providers
53 models · set one key, or several and switch| Provider | Environment variable | Models |
|---|---|---|
| DeepSeek | DEEPSEEK_API_KEY | V4 Flash, V4 Pro, V4 Flash Vision (reads images) |
| Anthropic | ANTHROPIC_API_KEY | Sonnet 5, Opus 5, Opus 4.8, Fable 5.1, Fable 5, Haiku 4.5 |
| OpenAI | OPENAI_API_KEY | GPT-5.6 Terra, Sol, Luna |
| Gemini | GEMINI_API_KEY | 3.8 Flash, 3.7 Flash, 3.1 Pro, 3.5 Flash-Lite, 3.6 Flash, 3.5 Flash, 3.1 Flash-Lite |
| Qwen | DASHSCOPE_API_KEY | Qwen3.7 Plus, Qwen3.8 Max, Qwen3.7 Max, Qwen3.5 Flash |
| Kimi | MOONSHOT_API_KEY | K2.6, K3, K2.7 Code, K2.5 |
| GLM | ZAI_API_KEY | GLM-5.3, GLM-5.3 Flash (reads images), GLM-5.2, GLM-5, GLM-4.7, GLM-4.7 FlashX |
| xAI | XAI_API_KEY | Grok 4.6, Grok 4.5, Grok 4.3 |
| Mistral | MISTRAL_API_KEY | Medium 3.5, Large 3, Small 4, Ministral 8B |
| Groq | GROQ_API_KEY | Llama 3.3 70B, GPT-OSS 120B, and whatever else your key serves |
| Cerebras | CEREBRAS_API_KEY | GPT-OSS 120B, GLM 4.7, and whatever else your key serves |
| Meta | MODEL_API_KEY | Muse Spark 1.3, Muse Spark 1.3 (Contributor), Muse Spark 1.2, Muse Spark 1.2 (Contributor) |
| MiniMax | MINIMAX_API_KEY | M2.7, M3, M2 |
| Tencent | TOKENHUB_API_KEY | Hy3, Hy4 Preview |
Who's using it
anonymous, opt-out any timeWhere it stands
Mindweave is developed and tested on Windows, and that is what it supports today. The suite does not yet pass on macOS or Linux — it hangs partway rather than failing outright, which points at process handling that has only ever been exercised on one platform.
Mindweave 1 is out, and v2.3.0 is the current release. Every change is written down, including the things that were quietly wrong. Read the changelog →