There are a lot of AI coding tools now, and a suspicious number of them were vibe-coded over a long weekend. Minnow is the small fish that made it. Plug in Claude Code, Codex or any model and get the editor, git, issues, memory and multi-agent builds around it. No account. No credits.
A totally unbiased comparison.
| A typical AI harness | Minnow | |
|---|---|---|
| Price | $20 a month, then credits, then “fair use” | $0. It’s AGPL, so nobody can change that later. |
| Your Claude or ChatGPT plan | Ignored. Buy our credits instead. | Bring it. Claude Code, Codex and Cursor plug straight in. |
| Account | Sign in with Google to continue | What account? |
| Tools | 20 MCP servers and a prayer | Browser, search, git and terminals, built in |
| Long chats | Forgets your name around hour two | Compacts itself. Still remembers hour one. |
| Parallel agents | Edit the same file and fight over it | Each gets its own worktree. No turf wars. |
| Bugs it spots | Mentioned once, lost in the scrollback | Filed as real issues, then picked up next |
| Everything else | A chat box and a nice gradient | An editor, git client, issue tracker and scheduler |
Everything an agent reaches for ships in the box and is on from the start. Add MCP servers if you like. You won't need twenty.
Near the limit, Minnow folds older turns into a summary of the goal, the files, the commits and what's still broken. Nothing is deleted, and the agent can look up any earlier message word for word. So you just keep going.
⎯ 214 messages folded into “Prior context” · originals kept · recall_history on
You already pay Anthropic or OpenAI every month. You don't need to pay a middleman as well. Sign in once and Minnow runs on the plan you have, limits and all. Got a spare API key or a GPU? Hand them the chores.
Hand Boards a plan. It splits it into tasks, runs builders in parallel in their own worktrees, has a separate tester check each one, and merges only what passes. Hit a rate limit at 3 a.m.? It waits for the reset. Laptop rebooted? It picks up where it stopped.
WAVE 1
boids.jsmergedWAVE 2
WAVE 3
+212 −31The editor, terminal, git, issues and memory are already attached, and they all share the agent's context. Nothing to paste between windows.
Language servers, a real code editor, terminals and git, right next to the thread. Open what the agent changed, run the tests yourself, carry on.
Route sub-agents, completions and utility jobs to different models. The clever one thinks, the cheap one writes titles.
Tag a GitHub issue minnow and the watcher plans it, builds it on a board and opens the PR once the tests pass.
"I noticed three problems" becomes three real issues instead of a paragraph you scroll past. Then the agent picks up the next one.
Split a job and each sub-agent gets its own checkout and its own saved chat. Watch them from one thread, or switch away and let them get on with it.
They all run on one engine. What the agent learns in one of them, the others already know.
Language servers, inline completion, Quick Edit, terminals, dev servers and a real browser.
Changes, history, branches, stashes, worktrees. PRs and CI through your own gh.
List, pipeline and triage, with tools the agent uses to file its own work.
Subscriptions, keys and local models, with routing and cost. Tells you what your GPU can actually run.
Memory as a markdown wiki on your disk. The agent reads and writes it. You can fix it when it's wrong.
Agent jobs on a timer. Nightly test summaries, Monday dependency checks, run history.
A plan run as waves of Builder and Tester agents in isolated worktrees.
It's AGPL-3.0. No plugin store to get approved by, and no enterprise tier coming to take your favourite feature away.
$ git clone https://github.com/HenriGrimm/Minnow.git $ cd Minnow $ npm install && npm start
Free, open source, no account, no credits. Install it, plug in the plan you already pay for, and go do something else.
v0.2.1 · release notes