Run Claude Code, Codex CLI, opencode, Gemini CLI or any terminal AI tool side by side. Give them work in parallel, see who needs you, review their changes and merge the best result. It runs entirely on your own machine.
01One task, three agentsEach gets its own branch, folder and port
Sample data · illustration
Transcript of the animation.
One task, three agents (Each gets its own branch, folder and port). A new task is typed once: add a dark mode toggle to the settings page. Two Claude Code agents and one Codex agent are picked. Each starts on its own git branch in its own folder, with dependencies ready and its own dev server port.
Every agent on one screen (Real terminals, side by side). All three agents run as their real command-line tools in a grid of live terminals. Each one edits its own copy of the code, so they never overwrite each other, and token use is counted per agent.
Someone needs you (Waiting agents jump to the top and alert you). Agent B stops at a permission prompt. Cockpit spots it on the screen, moves Agent B to the top in amber, shows one waiting for you, and sends a desktop notification. One click approves it and the agent carries on.
Turns end in checkpoints (Roll back any agent with one click). When an agent finishes a real turn, Cockpit knows, and saves a checkpoint of its files, including new ones. Any checkpoint can be restored with one click, and the current state is saved first.
Review, then talk back (Comment on a diff line, send it to the agent). The review tab shows every change file by file. A click on a diff line opens a comment, and all comments go back to the agent in one message, so it fixes its own work.
Compare, merge the best (Conflicts predicted before you merge). The compare panel puts the attempts side by side: files, lines added and removed, and tokens. It warns that two agents edited the same file and checks for conflicts with main. The best attempt is merged cleanly and a pull request is opened; the rest are discarded.
Run your agents like a team (Local, git-native, agent-agnostic). Syntora Cockpit. Stop babysitting terminals; run your agents like a team. It runs on your own machine, uses the coding agents and logins you already have, and keeps everything in ordinary git.
A sample task, drawn as an illustration. The repository, branches and token counts are fictional.
50+
Features
across 8 areas
47
API endpoints
local server
~5,800
Lines of code
TypeScript and JavaScript
3
Releases
4 and 5 October 2026
Why it exists
Running several agents turns into chaos
AI coding agents are good enough that one developer can run several at once. Without a cockpit, that means lost prompts, collisions and a review queue nobody can see.
Problem
Five terminal windows, no overview
Cockpit
Every agent in one dashboard, as its real CLI in a real terminal, with a grid to watch them all.
Problem
Agents step on each other
Cockpit
Each agent can get its own git branch and folder, so parallel work never collides.
Problem
Every parallel copy starts broken
Cockpit
Dependencies shared, .env copied, a setup script run, and a dev-server port of its own per agent.
Problem
Reviewing is the real bottleneck
Cockpit
File-by-file diffs, comments sent back to the agent, then merge, apply, open a PR or discard.
Problem
No idea what it costs
Cockpit
Tokens per agent (input, cached, output) and a running total, read from each agent's own transcripts.
How it works
A local server between your browser and your agents
Nothing goes through a third-party cloud. Agents use your existing CLI logins and subscriptions.
Headless terminal mirror that spots “waiting for input”
Claude Code hooks per session for exact turn and permission events
Git worktrees, checkpoints, merge, sync and pull requests
Token usage read from each agent's own transcripts
each agent on its own branch
Claude Codecockpit/claude-…
Codex CLIcockpit/codex-…
opencodecockpit/opencode-…
Gemini CLIcockpit/gemini-…
any CLIyour main checkout
50+ features
Eight areas, one screen
The strongest parts of each area. Every agent runs as its real CLI, unchanged.
Agents and terminals
Multi-agentClaude Code, Codex CLI, opencode and Gemini CLI detected automatically
Custom agentsAdd any terminal tool with its command and arguments
Broadcast and queueOne prompt to all or some agents, queued until each is free
Reload-safeClose the tab and agents keep running; tmux mode survives a restart
Attention
Needs-input detectionPermission prompts and questions spotted on screen, plus Claude Code hook signals
Attention-first listWaiting agents jump to the top in amber
Turn-finished detectionKnows a real turn ended, not just a pause in output
AlertsDesktop notifications, optional phone push and webhooks
Tasks and best-of-N
New taskOne prompt to one or many agents, started at once
Best-of-NThe same task on several agents, then compare
SynthesizeA fresh agent reads every attempt and builds a combined solution
LeaderboardWhich agent's attempts get merged most often
Isolation
Branch per agentIts own checkout on its own branch
Ready-to-run copiesShared dependencies, copied .env and local files
Setup and run scriptsA unique port per agent (4600, 4610, ...)
In-app previewOpen each agent's running app inside Cockpit
Review and merge
Review tabFile-by-file diff, new files included
Inline commentsClick a diff line, write feedback, send it to the agent
Conflict predictionWarns before merging which files would conflict
Safe merges and PRsAborts cleanly on conflict; push and open a pull request
Checkpoints
AutomaticA snapshot after every turn, untracked files included
One-click restoreRoll back to any checkpoint; the current state is saved first
ManualA named snapshot whenever you want
Workspace
Explorer, search, editorQuick edits with syntax highlighting, also on an agent's version of a file
Git panelBranch, changed files and recent commits
Token usage and system monitorPer agent tokens, live CPU, RAM, disk and network
Shortcuts and themesKeyboard-first, light and dark
Security and privacy
Local onlyThe server binds to 127.0.0.1 and rejects other websites
No cloud relayCode and prompts stay on your machine
Scoped file accessEditor and file APIs cannot leave the project folder
Your config untouchedHooks injected per session with a per-session secret
Why it helps
Before, and with Cockpit
Before
With Cockpit
Juggling terminal windows and tabs
✓ One screen, every agent visible in a grid
Finding an agent stalled on a prompt long ago
✓ An amber alert, on the desktop or the phone
Agents overwriting each other's files
✓ Each agent on its own branch and folder
Re-installing deps and fixing port clashes
✓ Setup script, shared deps, a port per agent
Reading raw git diff and copy-pasting feedback
✓ Visual diff, click to comment, one-click send
Guessing whether a merge will conflict
✓ Conflict prediction, or ask the agent to resolve it
Fear of an agent wrecking work
✓ Automatic checkpoints with one-click restore
Picking one agent and hoping
✓ Best-of-N: compare, merge the best, or synthesize
A typical task
From prompt to merge in five steps
01
Open a project
Then click New task.
02
Describe the change
Pick agents, for example two Claude Code and one Codex.
03
Do something else
You get a ping when an agent needs you or finishes.
04
Compare
Review the diffs and leave inline comments.
05
Merge the best
Or synthesize the best parts, then discard the rest.
Honest scope
What it is, and what it is not
It is
✓A local control room for the AI coding CLIs you already use
✓Agent-agnostic: any CLI that runs in a terminal can be added
✓Git-native: branches, merges and pull requests are ordinary git
It is not
–Not an AI model: it uses your existing agents and their subscriptions or keys
–Not a cloud service: it runs on your computer at localhost
–Not a sandbox: agents run with your user's permissions, as in your own terminal
–Not multi-user yet: a single-developer tool today
–Not a replacement for your editor: the built-in one is for quick edits and review
–Built and tested on macOS; Linux should work, Windows is untested
How it was built
Three releases in two days
About 5,800 lines of TypeScript and JavaScript and 47 API endpoints, each release tested end to end with real Claude Code and Codex sessions.
4 Oct 2026
v1 · Core
Local server with real terminals, a React UI with live terminals, grid view, broadcast prompt, workspaces, explorer, search, git panel and system monitor.
5 Oct 2026
v2 · Isolation
A git worktree per agent, review with merge, apply and discard, compare panel, resume after restart, code editor, notifications and interrupt.
5 Oct 2026
v3 · Research-driven
Needs-input detection, best-of-N and synthesize, setup and run scripts with ports and preview, inline review comments, conflict prediction, PRs, checkpoints, token usage, phone push, webhooks and tmux mode.
What comes next
Roadmap
Near term
Agent-to-agent context over MCP
Blind judging during review
Cost in currency, with budgets and an automatic stop
A one-click macOS app
Start tasks from GitHub Issues or Linear
Mid term
Scheduled and event triggers
CI status per agent branch, merge only when green
Container sandbox mode
Agents on remote machines over SSH
Tasks across several repositories
Long term
A mobile companion for approvals and review
Team mode with a shared dashboard
First-class Windows and Linux
Analytics on which agent wins at which task
Built with
Tech stack
Node.js
Express
WebSocket
node-pty
xterm.js
React
TypeScript
Vite
CodeMirror 6
git
tmux (optional)
We build tools like this for teams
Cockpit started as our own answer to running several coding agents at once. If your team needs internal tooling around AI, we build that.