SwarmPane/The product/34 features
Agents write fast. You decide what ships.
SwarmPane runs Claude Code, Codex and the other coding CLIs you use side by side on your computer, checks their work against your tests and keeps every step reversible. Here is every feature, by the job it does.
SwarmPaneAt a glance
- Runs
- Claude Code, Codex and the other coding CLIs you use
- At once
- Up to 16 agents, in real terminals
- Judged by
- Your own tests, typecheck and lint
- Undo
- Every agent step, byte for byte
- Where
- On your computer, on the accounts you already have
- Ships
- Nothing, until you click
Read what happens
Nine terminal windows float apart: Claude Code, Codex, Gemini CLI, Grok Build and plain shells, with notifications piling up. They fold into the SwarmPane mark, which opens into the app’s Code mode with four agents in a 2×2 Agent Grid. Broadcast sends one prompt, “Block cancelled plans from Pro routes”, to all four. Each edits the route and runs the tests: Claude Code, Grok Build and Gemini CLI pass 21 of 21; Codex fails two and asks before editing the tests, so its pane waits for you and the Dashboard lists it first. An Agent Race lets the tests pick between Claude Code and Codex, and nothing merges until you click. An agent step is rewound byte for byte, leaving a file you edited since. Overnight Autopilot works through three tasks while you sleep; by morning two are ready and one is marked failed. Up to 16 agents fill the grid, fold back into the mark, and the film closes on SwarmPane: Agents write fast. You decide what ships.
01/07Code with agents10 features
Agents in five terminals? Put them side by side.
Code mode runs the CLIs you already use in real terminals over your own folders, with files, Git and a browser beside them. Agent and Chat modes give the same agents a teammate’s memory or a quiet place to plan.
Code mode
Up to 16 agents in real terminals, each pane showing whether it’s working, waiting for you or done.
Status comes from each CLI’s own signals: Claude Code hooks, terminal notifications and exit codes. A CLI that reports nothing says “limited status” instead of guessing.
Pick a grid from a picture: 1 to 16 panes, or two beside the Browser.
A layout that can’t fit your window stays in the menu and says why. A launch set opens several agents in one click, each in its own worktree if you want.
Which agent needs you? Every session, the ones waiting first.
Needs you, failed or stuck, working, quiet, idle and done, across every project. Open a row to jump to its pane; Close idle tidies up the rest.
What every agent is doing, in a sidebar that stays open in every mode.
Each pane in each project, most urgent first, from the same live status the grid reads.
One prompt for several agents? Send it once.
Type a prompt or a command, pick the panes, and Broadcast sends it to each and says how many took it.
Your running app in the sidebar, kept open while you switch modes.
A preview gets its own cookies and storage, separate from the app, and it is where Point and Fix starts.
Building for iPhone? Keep a Simulator beside the code.
Boot a device, watch its screen in the pane, and open the full Simulator when you need it. Needs a Mac with Xcode.
Check the change without leaving the panes.
Read and edit files, review the diff, stage and commit, open a worktree, and search the whole project.
Agent and Chat
Agents with a job, instructions, skills and long-term memory.
Each turn runs on Claude Code or Codex, and a teammate asks before it edits your files.
Think it through before anything runs.
Plan with files and images as context, switch between Claude Code and Codex as you go, and keep every conversation.
Claude Code
~/launchpad
❯ Fix the failing checkout test
Read src/checkout/total.ts
OpenAI Codex
~/api
❯ Update the rate limiter and run the tests
exec edit src/limits.ts +6 −2
Gemini CLI
~/launchpad
❯ Write the release notes draft
Wrote docs/release-notes.md (34 lines)
~/launchpad % npm run dev
VITE ready in 412 ms
➜ Local: http://localhost:5173/
02/07Trust the result5 features
The agent says it’s done. Your checks decide.
Your own tests, typecheck and lint judge the work, every agent step can be rewound exactly, and nothing touches your files or your branch without your click.
Two agents race. Your tests pick the winner.
Each works in its own copy of the project, your checks rank them, and nothing merges without your click. Before it starts, a race says it costs about one run per agent.
See a bug in your app? Click it and say the fix.
SwarmPane finds the source, makes the change in a copy, runs your checks and shows before and after. Keep it, undo it or refine it.
Undo any agent step, byte for byte.
A checkpoint before every step that can edit your files. A rewind refuses to overwrite edits you made since, and says which.
Beta
Hand over the goal and the checks. Get back a checked result.
Jev picks one of your agents, runs your checks, and when one fails hands the work to another agent with the evidence. You review before accepting. Starter includes 5 million Jev tokens, once; Pro includes 10 million.
Nothing edits your files until you approve it.
Each request is signed on your computer, so one that changed after you saw it is refused. Stop ends an approved run and everything it started.
Baseline on d0f4c4a: test ✗ (3 failing) · typecheck ✓ · lint ✓
03/07Run unattended6 features
Can’t sit and watch? Come back to checked work.
Queue work for the night, chain agents into a process, or put a job on a schedule. Each one reports back with what passed and what didn’t.
Beta
Queue it tonight. Review it at breakfast.
Each task runs alone in a copy of the project and counts as ready only if your checks pass. Nothing is pushed without your approval, never to your default branch, and the morning report comes with a replay.
Labs, off by default
Have to close the lid? The agents keep working.
Your laptop stays awake with the lid closed only while an agent works, and sleeps as soon as none does. It also sleeps at the battery level you pick, when it runs hot, after the longest time you set or after 10 minutes offline, and the lid opens to a summary of what happened. Mac and Linux.
A process you repeat? Chain the agents once.
Up to 16 steps. Each result moves to the next agent, approval steps wait for you, and every run is saved.
Recurring work, on a schedule.
Run a teammate on an interval, every day or every week while the app is open.
Track the outcome, then hand it to a teammate.
A task runs with an agent and comes back for your review. Agents and MCP clients can create tasks too.
Labs, off by default
A board that moves with your real runs.
Races, Autopilot tasks and panes move their own cards, and only you mark a card done. It ships off while it’s finished.
Overnight AutopilotBetaRunning
Sat 3 Oct · launchpad
Running · 0 of 3 ready · keeping this Mac awake while it’s plugged in
- Ran
- 01:10–…
- Cut-offs
- 06:30 · 07:00no new attempts · all work stops
- Base
- main fbbf6ecyour working tree is never touched
04/07Cost and limits4 features
Usage limit at the worst moment? Keep going, on budget.
SwarmPane counts only what your CLIs report, stops runs at the budgets you set, and turns a usage limit into a choice instead of a dead end.
Token Guardian
Runaway run? It stops at your budget.
Only usage your CLIs report is counted, never estimated. Daily budgets stop runs before they overspend (chat is warned, never blocked), and a run that loops on one failure is stopped with the evidence.
Limit handoff
Hit a usage limit? Keep going.
Switch to your other account, or from Claude Code to Codex and back, with a brief of the work you edit first. Or resume after the reset, even if your computer slept.
Usage
What did each CLI actually use?
One row per CLI from its own reports, estimates labelled as estimates, and Codex’s live limits when Codex shares them.
Accounts
Work and personal sign-ins, side by side.
Save several Claude Code and Codex accounts, set a default, and pick the account for each pane or conversation.
05/07Voice2 features
Too much to type? Say it.
⌘⇧Space (Ctrl+Shift+Space on Windows) puts your words in the composer or the terminal as a draft. SwarmPane Voice comes with Pro, with 1,500 minutes of cloud voice a month; Whisper also runs offline on your Mac. SwarmPane Voice works on Mac and Windows, with Linux next.
SwarmPane VoicePro
Speak the prompt; edit the draft; send it yourself.
In a terminal the words are typed and never run. Audio is never stored: SwarmPane cloud transcribes and drops it, and Whisper keeps it on your Mac. See SwarmPane Voice
Voice AnywhereLabs, off by defaultPro
Dictate into any Mac app.
The same keys type into the app in front, never into password fields and never with a Return. ⌥⇧⌘Space speaks commands to SwarmPane, and anything that spends needs a click.
06/07Video1 feature
Need a launch video? Describe it and watch it build.
An agent writes the video from your project’s own brand. Point at any frame to change it, then render an MP4 on your computer.
Video StudioLabs, off by defaultPro
From a sentence to an MP4, on your computer.
Plan, storyboard, check and render in one place, with captions and a score made in code. Turn it on in Settings → Labs. Verified on macOS so far.
07/07Connect and extend6 features
Working in other tools too? Connect them.
Share SwarmPane’s tools with any MCP client, give your agents more of their own, and make the app look the way you work.
Let your other agents use SwarmPane.
Claude Code, Cursor or any MCP client can list teammates, create tasks, run routines and search your projects, over a local endpoint that needs a per-install token.
More tools for your agents, on your terms.
Add a trusted command to run on this computer, verify it, and see its tools before an agent uses them.
Teach it once, reuse it everywhere.
Reusable instructions, included whenever you assign the skill to a teammate.
Preview
Check your service CLIs’ sign-ins in one place.
Verify the GitHub, GitLab, Stripe, Cloudflare, Supabase, Vercel and Shopify CLIs you’re signed in to. Actions through plugins aren’t built yet.
Preview
Keep your servers’ host keys in check.
Save a server, verify its host-key fingerprint and test the connection. Running agents on a remote machine isn’t built yet.
Make it yours. Keep it readable.
Ten themes with matching terminal and editor colours, and a custom accent that keeps text contrast in check.
Connect a coding agent
Terminal command · Adds SwarmPane for every project on this computer.
claude mcp add --transport http --scope user swarmpane http://127.0.0.1:4192/mcp --header "Authorization: Bearer omcp_••••••••••••a7Kq"
Added it before? Run "claude mcp remove swarmpane --scope user" first. An entry named "overmcp" is this server under its old name: remove it the same way.
Tools your agents get
agentra_list_agentsagentra_list_tasksagentra_create_taskagentra_list_routinesagentra_run_routineagentra_searchagentra_search_project
The access token is per install and hidden.
→StartStarter or Pro
Bring your agents. Keep what passes.
Every feature on this page is in Pro. Starter has all of them except SwarmPane Voice and Video Studio. Your agents run on the accounts you already have.
- Starter
- 31 of 34 features
- Everything except SwarmPane Voice, Voice Anywhere and Video Studio
- Pro
- 34 of 34 features
- Adds SwarmPane Voice, Voice Anywhere and Video Studio