CPilot

A superlight native macOS app for running AI coding sessions from your menu bar. Tabs for many sessions, one calm chat, approvals you control, and pull requests from branch to merge.

Install with an AI agent

Paste this into an AI coding agent such as Claude Code. CPilot detects and installs its own prerequisites on first launch, so the agent does not have to:

Install CPilot on this Mac.
1. Run: curl -fsSL https://raw.githubusercontent.com/kalahroudi/cpilot/main/scripts/install.sh | bash
2. CPilot opens and its first-run setup checks what it needs and installs whatever is missing. Tell me to click
   "Set everything up" in that window if it is waiting, and wait until it reports everything is ready.
   If a step needs my password, tell me exactly which command to run in Terminal.
3. Register its MCP server: claude mcp add cpilot --scope user -- /Applications/CPilot.app/Contents/MacOS/cpilot-mcp
4. Call get_status and tell me what it reports.
Never call a destructive tool without asking me first.

Or by hand: download the latest disk image from Releases, or run the install script above.

What it does

Sessions

• Many sessions, one pane. Sessions are tabs or a side list, for the repository chosen in one switcher at the top of the menu. Only the open tab has a live view; idle tabs stop their agent and resume later with full context. Option-Command-N starts a quick session: Claude on Sonnet in its own worktree on a new branch. The mouse's back and forward buttons, a two-finger swipe and Command-[ and Command-] retrace the panes and sessions you visited.
• One calm chat. Replies stream as Markdown with highlighted code, sub-agent runs nested under the step that started them, and edit previews as diffs. Drop files or images anywhere to attach them. Messages you send while the agent works are queued and go out when the turn ends. After each turn Claude drafts your likely next message, faint in the box; Tab fills it in. Open a session you left and a recap says where things stand.
• Run what the agent tells you to run. Commands in a reply become Run chips that open an integrated terminal in the session's worktree (or Ghostty). Pages, SVGs, PDFs and images the agent writes can be rendered in the app, and every file a session wrote or changed is listed with a preview, its changes and a way to show it in Finder.
• Claude and other agents. Claude runs through its headless command line. Any agent that speaks the Agent Client Protocol can be added from the official registry in one click, including Antigravity, Cursor, goose, OpenCode and DeepSeek. Test on each agent reports whether it works, needs sign-in or what failed.
• Access you control. Read only, Ask first, Edit files or Full access, with per-project rules. Every action that needs approval shows a card with a short headline ("Edited Package.swift") and the details behind it: allow once, for the session, or always in this project. Full access needs Touch ID. Every decision is logged.
• Nothing lost on a crash. Messages are written to disk as they complete, streaming replies are checkpointed, drafts are saved as you type, and sessions resume by id with the agent's own context.
• Capture. Press Control-Option-Command-C to grab part of the screen, or add selected text, the front browser page or the clipboard. It lands in the active session's composer.

Shipping

• Pull requests, end to end. Create (CPilot brings the branch up to date with its base first, or hands conflicts to the agent), read the diff and commits, review with an agent, run a security scan, and merge with a box that works like GitHub's: merge styles, a bypass when rules allow it, a refusal if the branch moved, and a warning when the base already has the same change. Merge automatically when your rules pass. Open pull requests refresh while you watch, you are notified when checks pass or fail or a pull request starts conflicting, and an inbox lists every pull request that needs you. Everything goes through your own GitHub login.
• Stacks. Split a big change into small pull requests that build on each other. CPilot restacks when something changes underneath (including after a squash merge), submits the chain with the right bases, merges from the bottom, and offers an agent for conflicts. An agent can propose how to split a large branch. When pull requests in a stack edit the same files, one button combines them into a single pull request: Claude merges the pieces, removes duplicated code, checks it builds and passes, and CPilot opens the new pull request and closes the old ones. It uses plain git and the GitHub CLI.
• Specs. Start from a GitHub issue, a rough requirement you type, a conversation, existing code, a capture or an existing spec folder. A read-only agent drafts spec.md, plan.md and tasks.md; you approve each document, or let it run on its own and carry on after a restart. Each task then runs as a session on its own branch, stacked on the tasks it needs, and ticks itself off when its pull request merges. Specs live in a hidden .cpilot folder unless you choose to track them in git, and finished ones are archived.
• Repo setup in one go. README, license, GitHub Pages, a protected main branch and a scan on every pull request.

Anywhere

This is the beginning of the README. Read the rest at the original. Read the full README ↗