Ask once
Describe what you need in plain words. One message can hold several unrelated tasks.
Fox is an open-source agent harness that turns your AI subscriptions and models into one team. Ask once. A lead agent plans the work and runs the right agents in parallel, right in your terminal.
Describe what you need in plain words. One message can hold several unrelated tasks.
A lead agent reads your request, finds the files, and splits the work into tasks.
Each task goes to the agent best suited for it. They run side by side, each on its own files.
The lead checks the work, fixes what failed, and sums up what changed.
Fox works with 12 API models, seven of them on free tiers, and runs the Claude Code and Codex subscriptions you already pay for as teammates. Pick a model per session, and let the lead hand each task to the worker that fits it.
One strong model plans and checks. Cheaper models do delegated work in the background, at the same time.
Write tasks claim their files up front, all or nothing. Nobody edits another agent’s work.
Every task ends done, failed, cancelled or escalated, and the lead picks up what’s left.
Agents in a session share one memory, and chats in a project can share it too. A side chat starts out knowing what the main one learned.
A text-only model gets its page layout checked as text: overlaps, cut-off text, low contrast. Set a helper and it can ask a model that sees.
Stop a single delegated task from the app. The others keep running, and the stopped worker reports what it finished.
Every agent reads files and runs commands. These are the parts Fox was built for.
Fox drives its own Chrome and reads pages as text, so even a model that can’t see images can work the web and check its own layout. When a page asks for a password, a card, a one-time code, a CAPTCHA or access to your account, Fox brings the window up with that exact spot ringed, and waits for you.
Claude Code and Codex run as workers inside Fox, on your own subscriptions, next to its API models. Write one line per model in roster.json and the lead delegates by it: who codes, who researches, who sits on the bench.
{
"default_worker": "deepseek-flash",
"notes": {
"claude-code": "give it every coding job",
"codex": "give it every data job",
"nemotron-super": "short briefs, a few searches"
},
"off": ["deepseek-pro"]
}
Every session runs in its own worker process and writes everything to an append-only log. If a worker dies mid-run, the hub resumes the chat from that log and tells the agent which tasks were cut off and what changed in the repo.
✗ worker exited session fox-landing ↻ resuming the chat from the log [Fox resumed this chat] interrupted: the worker stopped before this task finished · deploy git diff --stat: site/index.html | 142 ++++++++ site/styles.css | 64 ++++ ✓ the agent carries on from the deploy
Before Fox first touches a file, it keeps the original. The diff view shows exactly what Fox changed in each file, with added and removed lines, whether or not the folder is a git repo.
Ask how something works and Fox can sketch it. Flowcharts and sequence diagrams are drawn by Fox’s own Mermaid renderer, right in the terminal, on a board beside the chat.
┌───────┐ ┌───────┐ ┌────────┐
│ TUI │───▶│ hub │───▶│ worker │
└───────┘ └───────┘ └───┬────┘
│
┌────▼────┐
│ model │
└─────────┘
Open a folder in the tree.
Open source. Runs on your machine. Star the repo to get the launch.
★ STAR ON GITHUB