One workspace for every coding agent.
A native desktop app that turns the coding CLIs you already use into a team: a resizable terminal grid, isolated lanes, verification you can trust, and the controls to let agents work longer without watching them.
Up to sixteen panes, shaped by hand.
Every pane is a real terminal running a real CLI — Claude Code, Codex, Gemini CLI, Copilot, Cursor Agent, Grok Build, Aider, Kimi CLI, Droid, OpenCode or a plain shell. Pick a layout, then drag columns, rows and crossings until it fits the work. PaneScale remembers the shape per workspace.
Free cells keep a fully started session waiting at low priority, so a new pane is ready the moment you open it.
Hooks, terminal notifications and prompt detection mark every pane as working, waiting for you or done.
Send one prompt to every pane, with each pane’s name filled in — or split a task list, and every pane that finishes takes the next task. Busy panes and open questions are never interrupted.
Every agent in its own lane.
Turn on “Isolated worktree” when you add a pane and the agent works in its own Git worktree, on its own branch, from your current commit. Several agents can change the same repository at once without touching each other’s files — or yours. When one is done, you decide what happens to its work.
Created from your current commit. Your own working copy is never touched.
Review the full diff first, then merge by cherry-pick, fast-forward or apply only.
Merging asks for a one-time approval — and a conflict stops instead of guessing.
Every step has a way back.
Before each agent turn that can change files, PaneScale captures a byte-exact checkpoint under a private Git ref. HEAD, your index and your branches never move. Rewind a step or just some files; if you edited a file yourself in the meantime, PaneScale refuses to overwrite it.
Checkpoints are written with a temporary index, so line endings and filters can’t change your files.
Every file is compared with what the checkpoint expects before anything is written back.
Changed your mind? Redo is the exact inverse of the rewind.
Fix the UI by pointing at it.
The browser pane shows your local dev server through a loopback-only proxy. Hover to inspect, click to select, then say what should change. PaneScale locates the source, runs the fix as an agent turn in an isolated worktree that includes your uncommitted work, runs a quick check and shows the diff. Included in Pro.
React, Vue and Svelte debug info first, then a scored text search. Guesses are labeled as candidates.
Only GET requests to localhost, 127.0.0.1 or [::1] — never to the internet, never to PaneScale itself.
Apply through the approval-protected merge, keep it in the lane, refine it, or undo.
Spend on purpose.
PaneScale reads usage from each agent’s own event stream — Claude Code’s result events, Codex’s turn ends — and shows it per engine and per run. Set a daily budget for each engine and a run that crosses it is stopped. Loops are detected and paused before they get expensive.
Engines that don’t report usage are shown as “not reported” instead of an estimate.
Budgets block the next attempt and stop the current run — they are not just a color change.
Repeated identical work is flagged; mark a false alarm as “not a loop” with one click.
When one agent stops, the work doesn’t.
Choose a backup agent for overnight runs. If the lead gets stuck or runs out of attempts, the backup continues in the same worktree with the feedback from the last attempt. From any pane you can also hand off by hand: PaneScale writes the brief and drops it into the agent you pick. Included in Pro.
The changes stay where they are; the next agent inspects and keeps what is good.
The brief carries the task, the last failing check and what was tried.
Hand off between any two supported agents.
Queue it tonight. Review it at breakfast.
Fill a queue with your own tasks, “Tonight” cards from the board or GitHub issues via the gh CLI. A preflight checks that the night can start; then each task runs in its own worktree with one to five attempts, time, token and file limits, a loop guard and an optional backup agent. The machine stays awake while it works. Included in Pro.
Only with required checks passing, no regressions and no integrity warning.
Every number in the morning report comes from recorded rows, saved as JSON.
Review the diff, adopt with undo, and open a draft PR only by clicking — with a one-time approval.
Agents, tests and people — in order.
Start from a recipe — Feature, Bug, Release or Docs — or from scratch. Add up to sixteen steps: coding agent, model prompt via Ollama or your own endpoint, test command, browser check, Git review, MCP tool and approval. Reference earlier results by name with {{key}}, validate JSON, branch on conditions and run groups in parallel. Included in Pro.
Run manually, on an interval or whenever a new commit appears.
Step log, Git evidence, failure analysis and an exportable receipt.
Live status, approval cards, retry and stop for each run.
Product videos, directed by an agent.
Describe the video. A coding agent writes the plan, then one small canvas function per scene — and the stage shows every draft as it lands on a multi-track timeline. Pin notes to exact frames, compare before and after, or let three agents explore three directions side by side. Every edit is code, every change can be undone, and the preview is the render. Included in Pro as a Labs preview.
Twelve overlapping transitions, camera keys per scene and kinetic, plain or social-style captions with a readability meter.
Whoosh, impact, swell and more are generated on your machine — transitions get their sound automatically.
One timeline re-composes for 16:9, 9:16, 1:1 and 4:5. The storyboard shows each scene in all four.
MP4 (H.264 + AAC), WebM (VP9 + Opus) or a looping GIF — frame-exact, with motion blur, faster than real time. No ffmpeg.
Syntax, runtime, determinism, stillness, pacing, caption reading time and contrast — rolled into a score the agent can fix.
Logo reveal, kinetic title, browser and phone mockups with a clicking cursor, terminal, code, charts, quote and end card.
Talk to the grid.
Dictate into the focused terminal or any text field, or speak a command. Speech is transcribed locally with whisper.cpp — the model is downloaded once, checksum-verified, and the recording is deleted right after transcription.
Layouts, tonight’s queue, new tasks, stop, open any mode, “what needs me”.
Commands are understood in both languages.
No cloud transcription unless you choose it.
Everything around the grid.
Teammates, chat, a board and schedules — built into the same app, on the same projects.
Persistent agents with a role, rules, an engine, allowed folders and a long-term memory that is added to every prompt as context — never as instructions.
Streamed answers, attachments, @ files and plan-then-approve edits — with checkpoints and a built-in review.
Reusable instructions like review checklists or code conventions, given to agents as context.
Start a card on an agent, review the result with diff and undo, send it back with a note or queue it for tonight. Only you can mark a card done.
Prompts on an interval, daily or weekly, each run in a fresh thread. A run missed while the computer slept catches up once.
Agents can read the board, move cards, list workflows and notify you. Off by default, loopback only, bearer token.
Guardrails you don’t have to remember.
PaneScale assumes agents will sometimes be wrong, and makes being wrong cheap.
Agents only run in folders you approved. The gate switches on as soon as you approve the first one.
Operations like publishing are signed with HMAC-SHA256 and can be claimed exactly once.
Lanes branch from your current commit. Closing one asks: keep, merge or discard.
Every file-changing step is captured first, and rewinds never overwrite your own edits.
Agent turns have timeouts, output caps and process-tree kills, so nothing runs forever.
Long runs keep Windows awake without forcing the display on — only while they run.
See it on your own repo.
Download PaneScale, open a folder and start a second pane. It takes about a minute.