Run claude, codex, antigravity and more in parallel, each in its own git worktree

Run claude, codex, antigravity, opencode and more in parallel. Each in its own git worktree.

Every new feature the day they ship it: the real CLIs on your own Pro / Max plan, no vendor backend.

Install · What it does · Sandbox · vs. Conductor · Contributing

Termic is a free, open-source desktop app that runs your AI coding-agent CLIs side by side, each isolated in its own git worktree, with an optional per-task macOS sandbox cage. It spawns the real claude, codex, agy (Antigravity), copilot and grok binaries (not the vendor SDKs), so inference rides on the Pro / Max plan you already pay for. Spin up four agents on the same branch, broadcast one prompt to all of them, watch a reliable work-done indicator tell you the moment each finishes.

Contributing: ideas first. I would love to hear what you would add to Termic. For a new feature, open an issue and let's discuss it before anyone writes code. To build something, pick an issue labelled planned. How contributing works.

---

Install

The recommended path is Homebrew + the official tap:

brew install --cask simion/termic/termic

That single command auto-taps simion/homebrew-termic, downloads the latest .dmg, and installs Termic.app into /Applications. Termic is signed with a Developer ID certificate and notarized by Apple, so there's no Gatekeeper warning.

Updates: Termic ships with a self-updater. When a new release lands you'll see an Update X.Y.Z pill in the top-right of the toolbar; click it to download + verify + relaunch. To check manually:

brew upgrade --cask termic

Direct download

.dmg, .app.tar.gz, and the ed25519 signature for each version live at the Releases page. The .dmg is signed and notarized by Apple, so it opens on first launch with nothing else to do.

Linux (AppImage)

Download termic_<version>_amd64.AppImage from the Releases page, make it executable, and run it:

chmod +x termic_*_amd64.AppImage
./termic_*_amd64.AppImage

The AppImage is ed25519-signed by the same CI flow as the macOS build, so the in-app updater works the same way: a new release appears as the Update X.Y.Z pill in the top-right, click to download + verify + relaunch. Keep the AppImage somewhere writable like ~/Applications/ so the updater can replace it in place.

The Seatbelt sandbox is macOS-only, so on Linux the task's Shield toggle is disabled. Docker mode is the Linux answer: it is not macOS-gated, so wherever Docker runs you can cage an agent in a container instead. Everything else (worktrees, parallel tabs, find-in-files, themes, in-app diff) works the same.

Wayland note: if fonts render thin, force X11 with GDK_BACKEND=x11 in front of the launch command (or in the .desktop file's Exec=).

Build from source

macOS (first-class)

git clone https://github.com/simion/termic
cd termic
make setup          # brew/rust/node + npm install + cargo check
make install        # build, copy to /Applications, launch

make dev (vite HMR + Rust auto-rebuild) is the iteration loop — see CONTRIBUTING.md if you plan to hack on the code.

Linux (build it yourself)

The signed AppImage on the Releases page is the recommended path for most users — see Linux (AppImage) above. Build from source if you want to hack on it, ship a .deb / .rpm for your own distro packaging, or run an unreleased commit.

bash scripts/setup-linux.sh (or make setup once make is installed) does all of the below on apt, dnf and pacman systems. By hand:

Prerequisites — Debian / Ubuntu (24.04+ has WebKitGTK 4.1):

sudo apt update
sudo apt install -y \
  build-essential curl wget file git pkg-config \
  libwebkit2gtk-4.1-dev libgtk-3-dev libayatana-appindicator3-dev \
  librsvg2-dev libssl-dev libsoup-3.0-dev libxdo-dev

# Rust stable
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y
. "$HOME/.cargo/env"

# Node 20+ — distro package, nvm, fnm, asdf, or mise (whichever you use)

Fedora:

sudo dnf install -y \
  @development-tools curl wget file git pkgconfig \
  webkit2gtk4.1-devel gtk3-devel libappindicator-gtk3-devel \
  librsvg2-devel openssl-devel libsoup3-devel libxdo-devel

Arch:

sudo pacman -S --needed base-devel curl wget file git pkgconf \
  webkit2gtk-4.1 gtk3 libayatana-appindicator librsvg openssl libsoup3 xdotool

Build and install:

git clone https://github.com/simion/termic
cd termic
npm install
npm run tauri build           # ~5 min first time, faster on incremental

The bundles land under src-tauri/target/release/bundle/. Pick whichever fits your distro:

# Debian / Ubuntu / Pop / Mint
sudo apt install ./src-tauri/target/release/bundle/deb/termic_*_amd64.deb

# Fedora / RHEL / openSUSE
sudo dnf install ./src-tauri/target/release/bundle/rpm/termic-*.x86_64.rpm

# Distro-agnostic — no install needed, just make it executable and run
chmod +x src-tauri/target/release/bundle/appimage/termic_*_amd64.AppImage
./src-tauri/target/release/bundle/appimage/termic_*_amd64.AppImage

After the .deb / .rpm install, "Termic" shows up in your application launcher. The in-app updater only knows how to replace the AppImage in place — .deb / .rpm users upgrade via git pull && npm run tauri build

• reinstall.

If the window looks slightly off — an empty gap on the left of the top bar, for example — that's the 84px reservation for macOS traffic-light controls. Harmless, will be cleaned up when the cross-platform chrome lands.

Wayland note: if fonts render thin, force X11 with GDK_BACKEND=x11 termic (or set it in the .desktop file's Exec=).

Windows (experimental)

Each release carries a Windows 11 x64 installer, Termic_<version>_x64-setup.exe, on the Releases page. It installs per user (no admin rights) and updates itself like the macOS and Linux builds. It is not code-signed yet, so SmartScreen asks first: More info, then Run anyway. Windows support is new and not yet used day to day. Docker mode is the sandbox on Windows; the macOS Seatbelt sandbox does not exist there.

To build it yourself, from Git Bash (it comes with Git for Windows):

git clone https://github.com/simion/termic
cd termic
bash scripts/setup-windows.sh   # installs build tools, Rust, Node, make via winget; npm install; cargo check
make install                    # builds the NSIS installer, installs it per user, launches it

Setup details, what differs from macOS, and what is not on Windows yet: docs/windows.md.

See CONTRIBUTING.md for the full dev guide.

---

What it does

Every agent runs inside a real PTY, the same binary you'd launch in iTerm, so there's nothing between you and the CLI. No vendor SDK (which bills against a separate credit pool as of June 2026), no metered markup, no backend daemon. Here's what the window gives you on top:

• Parallel worktrees. Each task is a git worktree under ~/termic/tasks/<project>/<name>/. Run N agents against the same branch across tabs; attach to repo root when you don't want a worktree; duplicate a worktree to spin up a parallel attempt off the same tip.
• Command palette (⌘K). Search and run any action — new task, file picker, find-in-files, rename, archive, YOLO, sandbox, theme, sidebars, settings — from one place, each with its shortcut inline. ⌘N is a quick project picker: fuzzy-find any repo and start a task without touching the sidebar.
• Multi-repo tasks. Group N repos (backend, frontend, infra) under one task with a shared CLAUDE.md and per-member dev-server ports. File finder, find-in-files, and the diff view span every member.
• Broadcast & Brainstorm. Send a single prompt to every agent in a task concurrently (⇧⌘B). Perfect for multi-agent code reviews, architectural brainstorming, or getting four "second opinions" on a complex bug in seconds.
• Config as Code (.termic.yaml). Persist all project-specific settings—setup scripts, run commands, preview URLs, and sandbox allowlists—into a .termic.yaml file. Commit it to your repo so your whole team gets the same optimized agent environment instantly.
• Per-task sandbox (macOS). Filesystem + network cage via sandbox-exec and an in-process HTTPS CONNECT proxy with a hostname allowlist. Lets the agent run with --dangerously-skip-permissions safely — the cage is the boundary, not the prompt. Or cage it in a container instead: Docker mode is an alternative isolation backend with a stronger filesystem boundary (experimental; network egress is not restricted the way the proxy restricts the seatbelt's).
• Sidebar Cockpit. Expand any task in the sidebar to see all its active agents, their live work-done indicators, and their agent-managed titles at a glance.
• Work-done indicator that's actually reliable. Per-CLI title classifier (Claude spinner, etc.) plus OSC 9;4, gated by byte-quiet and content-hash checks. This reliability enabled opt-in desktop notifications that only fire when an agent actually finishes a turn.
• Message Queues. Built on top of work-done detection: queue N messages (with optional repeats) to run autonomous "Ralph loop" sessions.
• Auto-Resume Everything. Termic auto-resumes sessions even for repo-root tasks, and resumes ALL agent tabs in a task, not just the primary one.
• Spotlight. Mirror one worktree's changes — committed, uncommitted, and untracked — into your repo root in real time, so your editor, dev server, and browser see the agent's work live. It runs on a detached HEAD and never commits to your branch; disabling it cleanly restores the checkout. (Conductor checkpoints your branch; Spotlight doesn't touch it.)

---

Sandbox

Optional per-task macOS Seatbelt (sandbox-exec) + an in-process HTTPS CONNECT proxy per task. Configured per project, pinned per task at creation (editable later from the task's Shield icon), enforced from the moment the agent spawns.

The cage:

• Writes restricted to the worktree, agent config dirs (~/.claude, ~/.codex), package caches (~/.npm, ~/.cache, ~/.cargo/registry), and TMPDIR. Always-denied: ~/.ssh, ~/.aws, ~/.gnupg, ~/.netrc, ~/.docker/config.json, ~/.kube. The macOS Keychain stays reachable unless the task turns on "Block Keychain access".
• Network restricted via an in-process CONNECT proxy with a regex hostname allowlist. Per-CLI vendor APIs (anthropic / google / openai)
  • GitHub + npmjs + PyPI + crates.io baked in. Add custom hosts per project. No external daemon — the proxy lives inside the Tauri binary, so there's nothing extra to install.
• YOLO auto-on inside the cage. The seatbelt profile IS the security boundary, so the agent's own permission prompts are skipped. The toolbar lightning icon turns red when YOLO is on without a sandbox (intentional danger signal — agents can rm -rf $HOME at that point).

Docker mode is the alternative backend: the agent runs inside a container rather than a seatbelt profile, for a stronger filesystem boundary and a blast radius that ends at the container. One generic image, editable as a plain Dockerfile in Settings, built by an explicit action; each agent's config folder is mounted separately so logins survive and cloning an agent keeps a work login apart from a personal one. Containers run as your host user, drop every Linux capability, and cap PID exhaustion. Marked experimental: network egress is not yet restricted the way the proxy restricts the seatbelt's.

For the full sandbox design — including the recent-denies debug panel and the auto-restart-on-edit flow — see CLAUDE.md §"Sandbox".

---

Status

• macOS: first-class — universal binary (Apple Silicon + Intel), signed updater, Homebrew cask. Requires macOS 12+ (Monterey).
• Linux: x86_64 AppImage shipped per release, signed by the same ed25519 key as the macOS build so the in-app updater works. ARM Linux + a Flathub submission are on the roadmap.
• Windows: experimental. x64 installer shipped per release, signed by the same ed25519 key so the in-app updater works; not code-signed yet. See docs/windows.md.
• Sandbox: the Seatbelt cage is macOS-only (sandbox-exec is Apple's frontend to it), so the Shield toggle is disabled on Linux and Windows. Docker mode is not macOS-gated, so a container cage is available wherever Docker runs.

---

Why use Termic over Conductor

The honest pitch — see termic.dev/vs/conductor for the full version with explanations.

	Termic	Conductor
License	Open source (AGPL-3.0)	Closed source, proprietary
Price	Free	Paid
Parallel agents in git worktrees	✓	✓
Attach an agent to the repo root (no worktree)	✓	✗ (always a worktree)
Runs claude	✓	✓
Runs codex	✓	✓
Bring your own agent (PTY-based)	✓ — opencode, aider, ollama, anything that runs in a terminal	✗
Multi-repo tasks	✓ — N repos under one wrapper, shared CLAUDE.md, per-member ports	✗
Sync a worktree into the repo root live	✓ — Spotlight, detached HEAD, never commits to your branch	◐ Checkpoints onto your branch
Command palette + fuzzy project / file switch	✓ — ⌘K / ⌘N / ⌘P	varies
Uses Claude Pro / Max subscription quota	✓ — spawns the interactive claude CLI directly	◐ Routes through the Claude Agent SDK
Monthly Claude cost on top of your Pro / Max plan	$0 — same quota as running claude in iTerm	Capped by the separate SDK credit ($20 / $100 / $200)
Local-only, no vendor backend in the loop	✓	✗ — vendor-hosted services
Per-task macOS sandbox (filesystem + network)	✓ — Seatbelt + in-process network allowlist	✗
Work-done indicator from real PTY signals	✓ — OSC 9;4 + per-CLI title classifier, no idle guessing	✗
Side-by-side ⇄ unified diff with syntax highlighting	✓	varies
Platforms	macOS, Linux (signed AppImage), Windows (experimental installer)	macOS

If you already pay for a Claude Pro / Max plan, Termic spawns the same claude binary that plan covers — no separate metered usage, no per-token markup. The agent and Anthropic still see the same auth they'd see in iTerm.

---

Contributing

Ideas first, code second. I am genuinely curious what you would add to Termic, and your idea is the most valuable thing you can bring. Writing the code is optional: spend your own tokens on it if you want to, or leave it to me.

• An idea for a new feature? Open an issue with the problem it solves and how you would use it. We talk it through there and decide together whether it fits Termic. Please do this before writing any code.
• Want to build something? Pick an issue labelled planned: those are agreed and ready. Comment /assign on it to have it assigned to you, then open the pull request.
• Small things go straight to a pull request. A bug fix, a typo, a small polish to something that already exists: no issue needed.
• A large new feature that arrives as a pull request, with no issue agreed first, gets parked. I will ask you to open the issue so we can discuss it, however good the code is. A feature that does not fit Termic costs your time and tokens and my review, and a conversation first saves both.

Agent-written pull requests are welcome on exactly the same terms. The dev setup, the pull request checklist and what does and does not fit the project are in CONTRIBUTING.md.

Roadmap

Two lists, and the difference between them is a commitment.

Planned is what will be built. Each one has an issue labelled planned, and that issue is where its progress lives.

Ideas are not committed to. Most have a write-up under docs/ideas/ arguing the case, and several are detailed enough to build from, but none of them is decided. They deliberately have no tracking issue: an issue implies someone intends to do it. Open one, or comment on the discussion linked below, if you want to make the case or pick something up.

Approved ideas move to docs/plans/ as implementation-ready specs and get an issue at the same time. That is the whole promotion path: docs/ideas/ → docs/plans/ + a planned issue → built.

Planned

• Mobile app. (#165) A companion app for checking on and steering tasks while away from the Mac.
• Apple Containers as a sandbox backend. (#296) A third cage beside the Seatbelt profile and Docker mode, using macOS's own container runtime, so a sandboxed task on Apple Silicon needs no Docker Desktop.
• Linear integration. GitHub and GitLab already ship: start a task from an issue, open the PR from the app, watch its checks and reviews from the right panel. Linear is the same shape and is not built.
• Wider Linux reach. ARM Linux builds and a Flathub submission. The Status section above already promises both.
• Opt-in usage telemetry. Anonymous, off by default, one toggle: which features are used and how often, plus crash reports. No code, no prompts, no file paths, no agent output, no project names.
• Termic in the official Homebrew cask repository. Today the install path is this repo's own tap, so it takes a brew tap first. Upstream would make it plain brew install --cask termic. The cask is ready (the .dmg is signed and notarized, and its livecheck makes it autobump-eligible), and Homebrew/homebrew-cask#274896 is filed and closed on one thing only: notability. Homebrew asks for 90 forks, 90 watchers or 225 stars when the author submits their own software. The forks already clear the third-party bar, so stars are the realistic lane, and at 225 the same PR gets reopened.

Ideas

• Windows, beyond experimental. An experimental Windows build ships in 1.9.0. What is left before it is supported: code signing, agent hooks for agents other than claude, a named-pipe control plane, and measurements on real machines. docs/ideas/windows.md.
• MCP server endpoint, phases B1 and B2. A scoped control plane an agent can call without being handed a terminal. Phase A (the stateless endpoint, the settings surface, one-click client setup) shipped in 1.0.0; the per-task bearer token that answers which task is calling and what it may do is what is left, and the 2026-07-28 spec revision reopened enough of the design that it is an idea again rather than a plan. docs/ideas/mcp.md.
• Intentional agent-driven orchestration. The plumbing already ships: an agent can spawn a task with --wait, prompt another, read its result and branch on the exit code. What is missing is intent, and an opinion about shape (fan out, queue behind, supervisor and workers). docs/ideas/agent-orchestration.md.
• Ambient agent status. A Dock tile, or a strip beside the Dock, showing what every agent is doing without bringing the window forward. docs/ideas/dock-widget.md.
• Fixing a language server that answers badly. Code intelligence ships, but when a server picks the wrong interpreter or indexes the wrong root there is nothing to turn. What the knobs should be, and which of them a user should ever see, is undecided. docs/ideas/lsp-tuning.md.
• Config sync, beyond manual. Syncing your setup through a private git repo you own ships as an experimental feature: syncing (pull and push) at launch, when a window comes forward, and a "Sync now" button. What is left: pushing on its own when something changes, and merging two machines' edits field by field instead of asking you to pick a whole file. docs/ideas/config-sync.md.
• The chip ships, and the questions it left open did not: what to show before an agent has taken its first turn, and whether the status line slot is the right thing to occupy at all. .

---

Sponsors

Termic is free, AGPL-3.0, and built by its author and a growing group of dedicated open-source contributors. If your team builds on AI coding agents and finds it useful, sponsoring helps keep it moving.

Also sponsoring: Vyttle, Sage Haven.

---

License

AGPL-3.0-or-later. Fork it, modify it, build a derivative — the only string is that derivatives stay AGPL too. The "open core that quietly went proprietary" pattern can't happen with this license, which is most of the point.

---

Links

• Website: termic.dev
• Issues: github.com/simion/termic/issues
• Releases: github.com/simion/termic/releases
• Homebrew tap: github.com/simion/homebrew-termic
• Contributing: CONTRIBUTING.md
• Architecture notes (for hackers + AI agents working in this repo): CLAUDE.md