Your ▲ ƒx agents on your own server, from any device.
fx is the coding agent from Vercel Labs. fox runs it on a VPS or a computer at home and gives you every session, every subagent and a terminal that never dies, from your phone or your laptop.
Independent project, not affiliated with Vercel. fx is vercel-labs/fx, Apache-2.0.
- Many fx sessions, one screen. Which one works, which one waits for you, what each subagent answered, what files they wrote.
- The real agents, nothing else. fox runs the
fxbinary, and optionally Claude Code, in tmux and reads what they write to disk. It never talks to a model, so it uses whatever you logged them into: a ChatGPT subscription for fx, your Claude Pro or Max subscription for Claude Code. Pick the agent per session. - Your server, your choice. Any VPS or a box at home. Agents keep working with every browser closed; you come back from any device, behind a passkey.
- 2026-11-03 Expose new fx settings as opt-in controls (#79)
- 2026-11-02 Expose new fx settings as opt-in controls (#76)
- 2026-10-31 Add clear all icon to replies inbox (#75)
- 2026-10-31 Clear all new replies at once from the reply inbox (#74)
- 2026-10-31 Live checks on what fx changed: lint and type errors next to its files (#72)
- 2026-10-31 Open a large tool result from its exact transcript call (#70)
All of it: CHANGELOG.md.
- Sessions grouped by workspace (with its jobs and open pull requests counted under it, one click to see only those), with a live dot (working, waiting, needs input), a bell when one finishes, and an inbox of replies you have not read.
- Subagents grouped by task, so the same question on three models is a side-by-side comparison; weak answers are flagged.
- Files fx wrote, opened the way they read best: pages and images rendered, Markdown formatted, CSV as a table, code highlighted. The newest page shows as a live card, sandboxed. Workspace settings can opt into checking changed files after each fx turn with the repo's
lint,typecheckorcheckscript (or a command you supply); errors stay in fox until you choose to insert them into fx's input. - One terminal per session in tmux; pin the ones you keep. Drop files on it, dictate into it (the text lands in fx's input; you send it). Toggle the read-only mosaic in the tab bar to watch up to four live sessions, then open a tile to type.
- Workspaces: a folder of GitHub repos, cloned once, with each repo's branch and changes at a glance; click the branch to go back to main, switch, update or start one (uncommitted work is kept in a wip/ branch, never dropped). Start a session in one repo and explicitly grant fx file access to selected sibling repos for that launch only.
- Jobs: scripts on a schedule, each with its model, log, summary and links. A job can have fx make a change and open a pull request it checked itself. docs/jobs.md
- Pull requests of your workspaces' repos in one list: what each does, its files, the review job's verdict; merge or close it from fox.
- Tools from Settings, for fx and Claude Code: a browser, Chrome DevTools, library docs, Vercel, Linear.
- Skills for fx in Settings: read discovered skills, preview and install GitHub skills globally or per workspace, and remove fox-installed skills. Click ✳ in the terminal bar to find skills for that folder and insert an unsent draft.
- Search across every session (⌥Space), metrics for the last 7 and 30 days, seven themes (one is a real star map).
- On a server: open it from any device, see the dev servers fx starts, and publish pages and tools as named apps (
https://<name>.<your domain>, private to you) that stay up after the session: a folder fox serves or a command fox keeps running. docs/server.md
| A server | A Linux VPS or a computer at home: Ubuntu 24.04, 4 GB of RAM or more. You use it from any device with a browser: Windows, macOS, Linux or a phone. |
| A model provider for fx | One of: a ChatGPT subscription (fx login codex), a Vercel AI Gateway key (pay per use; mixes models from several vendors) or a Grok subscription. |
| Claude Code (optional) | A Claude Pro or Max subscription. Install it with --claude-code and sign in once with claude auth login; then the + sheet lets you start a session on fx or on Claude Code. |
| A way in | A domain on Cloudflare (Tunnel + Access, free), or Tailscale if you have no domain. |
| A passkey | On the devices you will use. It is fox's login. |
| GitHub (optional) | gh with a fine-grained token, to open workspaces from your repos and let fx open pull requests. |
On your server (Ubuntu or Debian), one script sets up everything but your logins, which it tells you how to do. With a domain on Cloudflare, or --tailscale instead:
curl -fsSL https://raw.githubusercontent.com/jamarks/fox/release/setup-server.sh | sudo bash -s -- --cloudflare fox.example.comAdd --claude-code to also install Claude Code. What it does and the steps after it, logins included: docs/server.md.
To try it first on your own computer (macOS or Linux), one command checks Node, tmux and fx, installs what is missing, installs fox and opens it at http://127.0.0.1:4848, where you create a passkey:
curl -fsSL https://raw.githubusercontent.com/jamarks/fox/release/setup.sh | bashSettings and gotchas: docs/install.md. What fox reads and how: docs/how-it-works.md.
Option is the modifier, because ⌘T and ⌘W belong to the browser. ⌥? lists them all.
| Keys | Action | Keys | Action |
|---|---|---|---|
| ⌥T | New session | ⌥Space | Search all sessions |
| ⌥W | Close tab (fx keeps running) | ⌥\ | Show / hide the right column |
| ⌥[ ⌥] ⌥1…9 | Switch tabs | ⌥L | Next theme |
| ⌥P | Pin the tab | ⌥= ⌥- ⌥0 | Font size |
npm test runs everything with Node's own runner. Plain Node and browser JS, no build step. AGENTS.md is the map.
MIT. Built in an afternoon with fx 0.0.13 and a lot of tmux capture-pane. Issues and forks welcome.
