Plan a change, build it with agents, and review the diff in one desktop app. Your code, issues, and project notes give the agent the context to carry the work through.
Free and open source. No Minnow account. Local or cloud models.
INSIDE MINNOW A failing test, a one-file fix, and the diff ready to review.
Your code, issues, changes, and project knowledge belong in the same place. In Minnow, they’re built into the workspace your agent works in.
See how the work flowsThe implementation
The work ahead
The changes
The decisions
The same project. The same agent. Context that carries forward.
One conversation plans the work, runs it, shows you the diff, and remembers what it learned for next time.
Write what you want in plain language. No ticket template, and no other app to open first.
Plan mode has no tools that edit code, so nothing in your repo changes while you read. Adjust the plan, and approve it when it's right.
The plan becomes a board of Builder and Tester tasks, each on its own branch in a separate worktree, so your checkout stays clean. Their shell commands still run on your machine; turn on the agent sandbox for unattended runs.
Changes come back as a diff with the tests already run. Commit it, open a PR, or send it back.
Decisions go into Brain and bugs go into Issues, so the next run starts with that context.
Start in Code. Review in Source Control. Keep decisions in Brain. Explore the tools your agent can use without leaving the project.
File tree, LSP, inline completion, Quick Edit, terminal, dev servers, and a Chromium preview.
Chat works directly on the repo, so the agent edits the file you're looking at and shows you the diff first.
Review changes, history, branches, stashes, worktrees, pull requests, and CI.
The agent can commit and open PRs. You still decide what ships.
A plan becomes waves of Builder and Tester tasks.
Each task runs in its own worktree and reports progress to the board. Nothing merges until you approve it.
Hardware-fit scoring tells you whether a model will run well on your machine before you download it.
llama.cpp and MLX are built in, cloud providers use your own key, and routing assigns models to roles.
Brain is a Markdown knowledge base with semantic search, a code index, and memories the agent reads and writes.
Every page is a file on your disk. What one session works out, the next one starts with.
List, board, triage, and saved views, plus issue tools the agent can use.
If an overnight run hits an edge case, the issue is waiting for you in the morning with steps to reproduce.
Run agent jobs on an interval or a cron schedule, scoped to a workspace, with a history for every run.
Jobs run while Minnow is open and pause when you close it.
Cloud models, local models, or both. Choose how you connect.
No Minnow account. No Minnow subscription. Minnow puts no cap on usage. Cloud providers set their own prices and limits.
Every surface shares the same agent, tools, and models, so these work the same wherever you are in the app.
General, Build, Plan, and Debug. Each mode sets the prompt and decides which tools exist at all, so Plan mode can't edit your code.
Learn more →Hand a job to a sub-agent: explore the codebase, run shell commands, debug a failure, or write up an issue. It works in the background and reports back, so your conversation stays short.
Learn more →Every built-in tool can be set to Full, Ask, or Off. Each call shows up in the transcript as it happens.
Learn more →Reusable workflows like code review, opening a PR, and writing tests. Type / to run one. Each is a Markdown file you can edit or copy.
Learn more →Connect any MCP server or local plugin. Their tools follow the same permission rules as the built-in ones.
Learn more →The agent can map the repo, find symbols and their callers, and read LSP diagnostics instead of searching text blindly.
Learn more →Chats, settings, memories, and downloads live under ~/.minnow, as Markdown, JSON, and SQLite.
Nothing is locked in a format only Minnow can read.
Skills, permissions, rules, and prompts are plain text. Edit them in any editor and the next session picks up the change.
Mode prompts and Quick Edit prompts are files too, so you can read exactly what gets sent to the model and change it.
Minnow collects no telemetry and needs no Minnow account. API keys are encrypted on your disk. Cloud models receive the prompts and context sent to them.
~/.minnow is just a directory. Make it a repo and your prompts, skills, and settings get a history. Leave .key, providers/, search.json, and models/ out: they hold your credentials and gigabytes of model weights.
~/.minnow
├── .key
├── config.json
├── tools.json
├── sessions/
├── brain/
├── issues/
├── prompts/
├── skills/
├── providers/
└── models/
The point of a dotfile is that you can open it.
There's no Pro tier, no per-seat pricing, and no feature that unlocks when you pay.
Minnow is licensed under AGPL-3.0-or-later. Every release stays open source, and anyone who ships a modified version, even as a hosted service, has to publish their source too.
Minnow follows Blender's funding model: the full product is available to everyone, supported by people who use it.
Sponsoring supports the work. It doesn't unlock anything extra.
Available for Windows, macOS, and Linux. No account needed, and no API key if you use a local model. Minnow is in open beta, so expect frequent updates.
git clone https://github.com/HenriGrimm/Minnow.git