Run AI agents in parallel — each in its own git worktree. Bring your own API keys. Control the whole thing from your phone. Native to macOS.
Every agent moves through four states, bottom to top. When a PR lands, its slot frees and a fresh agent takes its place. The cycle never stops.
You hand a task to an agent on the host checkout. It reads, thinks, proposes — no branch yet, no isolation.
As soon as it commits, TermLoop spins up .termloop-worktrees/<branch>. Your host stays clean; the agent keeps iterating.
Push, open, wait for CI. The agent stays attached to its worktree and keeps responding to review without blocking anything else.
PR lands, the worktree is archived, and the agent card disappears — making room for the next task the moment you're ready.
You work in states, not commands. Four stops, one direction, forever upward.
TermLoop replaces your terminal. It doesn't bolt an AI panel onto VS Code. Built from the ground up for people who already live in tmux.
Run four Claudes, a Codex, and a Gemini at once — each in its own pane, each in its own worktree. Nothing collides. Nothing waits its turn.
Every task gets its own branch and directory, created automatically. No stashing, no branch gymnastics. Agents can't step on work in progress.
Start an agent from the couch. Check output from a coffee-shop queue. Native iOS client over an authenticated TCP bridge — no cloud round-trip.
You pay the model provider directly. No middleman markup, no rate-limit surprises. TermLoop never touches your API traffic.
Group workspaces by project, switch contexts in one keystroke. Your "work" project, your "side-project" project — each with its own agents and env.
Inspect the socket protocol, the agent loop, everything. No telemetry, no phone-home — what you see on your machine is all there is.
Three steps. No config files. No YAML.
Open TermLoop in any git project. It auto-detects remotes, branches, and dev server scripts.
Pick a model, give a task. TermLoop forks a worktree, boots the agent, and streams output to a new pane.
When the agent reports done, review the diff in place. Merge, discard, or hand off to another agent.
TermLoop is built on cmux — a GPL-3.0 terminal multiplexer by Manaflow. We publish our agent layer publicly too. No telemetry, no phone-home. Final license and pricing announced at launch.
None of those. TermLoop is built on cmux, an open-source terminal multiplexer — think tmux with a native macOS shell. We added a pane-per-agent model, git worktree orchestration, and a mobile bridge. Claude Code and the others run inside TermLoop panes, not the other way around.
Any model with a CLI: Claude Code (Anthropic), Codex (OpenAI), Gemini CLI, Aider, Cline, local Ollama, or your own scripts. TermLoop just hands each one a worktree and a pane — it's model-agnostic.
No. Agents call their providers directly from your machine with your own keys. TermLoop never proxies, stores, or phones home. The only network call it makes is checking for app updates, which you can disable.
Because we wanted a native, fast, low-latency UI before we wanted broad reach. Linux is on the roadmap. Windows isn't, for now.
A native iOS client connects to your TermLoop session over a password-authenticated TCP socket on your local network (or over a tailnet if you have one). Same protocol as the desktop — you can list workspaces, switch projects, and watch agent output live.
Early access is free. Final pricing will be announced before launch — expect a free tier for personal use and a paid tier for power features like the mobile bridge and cloud session sync. Sign up for early access and we'll tell you first.
TermLoop builds on cmux, which is GPL-3.0-or-later with a commercial option by its publisher. The full TermLoop licensing terms will be published alongside the first binary. Until then, the repository is source-available for inspection.
One email when TermLoop ships. Plus a private TestFlight invite when mobile opens up. No drip.