Built so coding agents earn their keep.
Watchfire is an open-source remote control for AI coding agents. We started building it because running these tools in the terminal felt like riding a motorcycle with no helmet — fast, sometimes useful, occasionally expensive.
What we're trying to do.
Coding agents are powerful, but the surface — a single terminal session — doesn't scale to multiple tasks, multiple projects, or multiple engineers. Verifying what an agent actually changed takes longer than the change.
Our bet is that every agent run should happen inside a sandboxed git worktree, the output should be a reviewable diff, and the operator should be able to see what was kept on disk in the time it takes to open a YAML file.
The shape that falls out: a daemon that orchestrates, thin clients (CLI/TUI and GUI) that observe, and a task model that's just a YAML file.
For the longer version — six tenets on how agents should run inside real codebases — read the manifesto.
How we make decisions.
Four rules that have survived contact with the codebase.
Boring on disk.
Plain YAML, plain git. If you can't read the state with cat and git log, we got it wrong.
The operator is the loop.
Watchfire doesn't decide when work is done — the operator does. We surface what matters and stay out of the way.
Six backends, one workflow.
We're agent-agnostic. The day a better coding agent ships is the day we support it.
Open by default.
Everything is on GitHub. No telemetry you didn't ask for. No tier above “all the features.”
Read more on the security page.
Who's building this.
Watchfire is a small open-source project. One maintainer today, plus contributors on GitHub.
Nuno Coração
Creator and maintainer
Nuno started Watchfire because he wanted to run coding agents without babysitting a terminal. He writes most of the daemon, the CLI/TUI, the GUI, and — in a properly recursive way — uses Watchfire to build Watchfire.
Plus contributors on GitHub. Want to help? Read the contributing guide.
Writing about Watchfire? Grab logos and boilerplate on the press kit.
How we got here.
The short version. The long version is the changelog.
Late 2025
First commit. A CLI that wrapped Claude Code and put each task in a git worktree.
v1.0.0 Ember (early 2026)
JSONL transcripts, a formatted conversation log viewer, and sandbox hardening for projects living anywhere on disk.
v2.0.0 Spark
Multi-agent support: Claude Code, OpenAI Codex, opencode, and Gemini CLI all become first-class backends.
v3.0.0 Blaze
GitHub Copilot CLI joins the lineup as the fifth first-class backend.
v4.0.0 Beacon
Two-way integrations, fleet and per-project insights, and the inline diff viewer.
v8.0.0 Inferno
The parallel-workspaces release: a multi-window GUI with one window per project, a chat-primary layout, wildfire mode in the GUI, a mission-control home window, and code-output analytics that measure what the agents actually shipped.
v9.0.0 Firestorm (today)
See the changelog for what's current.
Try it, star it, or come say hi.
Two minutes to install, one YAML file to your first task.
Read the docs, browse the blog, see what shaped our thinking on /resources, or come say hi on the community page.
Watchfire is open source — see /open-source for the license, contributors, and how to sponsor the project.