Run Claude, GPT, Gemini, Qwen and local models in parallel, each in the harness made for it, each on its own git worktree. Switch models without losing the conversation. Bring a friend into the session. Keep going from your phone.
Free and open source, MIT license. macOS 14 or later, Apple Silicon and Intel.
Drives the real agent CLIs
A model is only as good as the agent loop around it. Kabul runs each model in the harness built for it: Claude in Claude Code, GPT in Codex, Gemini in Antigravity or Gemini CLI, Qwen in Qwen Code, and anything else through OpenCode or Goose, over the Agent Client Protocol.
Local models are routed by family: a Qwen on LM Studio goes to Qwen Code, a gpt-oss goes to Codex. You pick the model; Kabul picks the harness, and tells you when it had to fall back.
Hit a rate limit, want a second opinion, or want a cheaper model for the boring part? Pick another model in the same chat, even one from a different provider or a different account.
Kabul keeps one canonical transcript and hands off to the new agent: the latest messages and who wrote them, the tools used, the diff of the branch and your .context/ notes. Same worktree, same branch, checkpoints keep going.
The cart totals test fails about one run in ten. Find out why.
The discount is rounded before the tax, so some totals drift by a cent. Fixed in cart/totals.ts.
Write regression tests for the rounding.
Added 3 tests at the rounding edges. 14 passing.
Reviewed the diff: zero-quantity items skipped the discount. Covered that too.
Invite someone into a workspace as editor or observer. You both see the agent stream live, send messages into one shared queue, approve tools (the first decision wins) and type in the same terminal, with presence and authorship on every message.
The same server puts every session of your computer on your phone, and pings you when an agent needs a decision.
Coding agents are good enough to work on several things at once. The tools around them are not: each agent lives in its own CLI, one task per terminal, tied to the desk, and locked into one vendor's models.
In Kabul each task gets its own workspace: a git worktree on its own branch, with its own ports, terminal and chats, so ten agents can work on ten things without stepping on each other. You review the diff, leave comments for the agent, watch the checks and ship the pull request in the same window.
.env* files copied in, statuses, pins and a board view.gh CLI, with multiple accounts. Routines run a prompt on a schedule or a PR event.Kabul drives the real CLIs, so you keep their behavior, their updates and your existing logins, settings and MCP servers.
| Agent | Talks over | Chat and tools | Approvals | Plan mode | Stop | Resume | Several accounts | Local models | Phone |
|---|---|---|---|---|---|---|---|---|---|
| Claude Code | stream-json | ✓Confirmed | ✓Confirmed | ✓Confirmednative | ✓Confirmed | ✓Confirmed | ✓Confirmed | ✓ConfirmedAnthropic API | ✓Confirmednative remote control |
| Codex | app-server | ✓Confirmed | ✓Confirmed | ✓Confirmednative | ✓Confirmed | ✓Confirmed | ✓Confirmed | ✓ConfirmedResponses API | ◐Partial |
| Antigravity | stream-json | ✓Confirmed | ✓Confirmedproject rules | ◐Partialrestarts to exit | ◐Partial | ✓Confirmed | ✓Confirmed | –Not available | ✓Confirmednative remote control |
| Gemini CLI | ACP | ✓Confirmed | ✓Confirmed | ✓Confirmednative | ✓Confirmed | ✓Confirmed | ✓Confirmed¹ | ◐PartialGemini-API gateways | ◐Partial |
| Qwen Code | ACP | ✓Confirmed | ✓Confirmed | ✓Confirmednative | ✓Confirmed | ✓Confirmed | ✓Confirmed | ✓ConfirmedOpenAI API | ◐Partial |
| OpenCode | ACP | ✓Confirmed | ✓Confirmed | ◐Partialemulated | ✓Confirmed | ✓Confirmed | ?Not confirmed yet | ✓ConfirmedOpenAI / Anthropic | ◐Partial |
| Goose | ACP | ✓Confirmed | ✓Confirmed | ◐Partialemulated | ✓Confirmed | ✓Confirmed | ✓Confirmed | ✓ConfirmedOpenAI API | ◐Partial |
| Cursor | ACP | ?Not confirmed yet | ?Not confirmed yet | ?Not confirmed yet | ?Not confirmed yet | ?Not confirmed yet | ?Not confirmed yet | –Not available | ?Not confirmed yet |
✓Confirmed◐Partial?Not confirmed yet–Not available
Confirmed means the installed CLI was run through Kabul. ¹ Isolated profiles work; signing in with a Google account through Kabul is not tested end to end yet. Any agent, and any shell, can also be followed and typed into from the phone through the live terminal mirror.
Agents and harnessesAn installable app for iOS and Android, served by your own computer over Tailscale. See what is running and what is waiting for you, approve tools, read the diff and the files, follow the live terminal, start a workspace or switch the model.
Push notifications tell you when an agent waits for a decision. iOS look on iPhone, Material 3 on Android, one codebase.
The phone app is a PWA you install from the browser. Store builds are on the roadmap.
.dmg and drag Kabul to Applications. You need git, tmux and the GitHub CLI: brew install git tmux gh.Only the visible workspace is mounted, events are batched per frame, long lists are windowed, and a concurrency limit queues agents with the focused workspace first. Measured budgets run in CI, so a regression fails the build.
Server load test (100 workspaces, 30 fake agents, 3 clients): 6.5% CPU on average, about 24 MB of memory, 0.17 ms from event to client at p95.
Methodology and full reports| Metric | Before | Now |
|---|---|---|
| Switching workspace (p95) | 1,098 ms | 48 ms |
| CPU with 30 agents emitting | 100% | 8% |
| Frames per second with 30 agents | 9.7 | 60 |
| Key press to frame (p95) | 178 ms | 19 ms |
| Opening a 2,000-message conversation | 1,530 ms | 68 ms |
| JS heap | 174 MB | 20 MB |
Synthetic loads on one Mac; real agents emit far fewer events.
Workspaces that run in the cloud, so your agents keep working with your Mac closed. Invite your team and reach every session from anywhere, without setting up Tailscale.
Kabul is built in the open. Issues and pull requests are welcome.
Small, well-scoped tasks to learn the codebase.
Good first issuesLayout, commands, commits, versions and how CI works.
CONTRIBUTING.mdThe app is in English and Portuguese. Every string lives in one JSON file per area, so a new language is a pull request.
Translation filesTell us what happened, the version and how to reproduce it.
Open an issuePlease report security problems privately, never in a public issue.
SECURITY.mdKabul is free and open source. Sponsorships pay for the time that goes into it.
Not sponsoring? A star on GitHub helps too. Star on GitHub
Yes, MIT licensed. You bring your own agent subscriptions or API keys (Claude, ChatGPT, Google, local models). Kabul doesn't resell or proxy model access.
Only what the agents send to their providers and what gh sends to GitHub, exactly as if you ran those CLIs yourself. Kabul has no server of its own and no telemetry.
No. It drives the real CLIs (Claude Code, Codex, Antigravity and others), so you keep their behavior, their updates and your existing logins, settings and MCP servers.
macOS 14 or later (Apple Silicon and Intel) is the main platform. Windows 10 1809+ and 11 are in beta: terminal sessions don't survive closing the app yet. Linux is not packaged. The phone app runs on iOS and Android.
Only for the phone, other machines and collaboration. The phone server listens on localhost and your tailnet, never on the open internet.
No. An invite is scoped to one workspace. But an editor can run commands as your user (terminal and approvals), so only invite as editor someone you would let use your computer.
Not yet. The phone app is a PWA you install from the browser; store builds are on the roadmap.
Yes. The desktop app, the phone app, the menus and the notifications are in English and Portuguese (Brazil). Kabul follows the system language, and you can pick one in Settings › General › Language.