MCP server
ffgit ships a Model Context Protocol server, ffgit-mcp, inside the app. Coding agents read repositories through ffgit’s own Git core and can ask the running app to change them. The server itself never writes.
Set it up
Section titled “Set it up”Main menu → Coding agents (MCP)… shows ready-to-copy commands that already allow the current workspace’s project folders. The helper is /Applications/ffgit.app/Contents/MacOS/ffgit-mcp on macOS, ffgit-mcp.exe beside ffgit.exe on Windows (%LocalAppData%\Programs\ffgit by default) and /usr/bin/ffgit-mcp on Linux (or ffgit-mcp in the tarball’s ffgit folder); pass one --allow per folder.
claude mcp add --scope user ffgit -- \ /Applications/ffgit.app/Contents/MacOS/ffgit-mcp --allow ~/codecodex mcp add ffgit -- \ /Applications/ffgit.app/Contents/MacOS/ffgit-mcp --allow ~/code{ "mcpServers": { "ffgit": { "command": "/Applications/ffgit.app/Contents/MacOS/ffgit-mcp", "args": ["--allow", "/Users/you/code"] } }}Agents you start from ffgit get these tools automatically for their session. With them a run can also report short progress — its phase, what is done and the next step — which its card shows as Agent reported; reports never change files or Git.
Read tools
Section titled “Read tools”Read-only, non-destructive and idempotent:
| Tool | What it returns |
|---|---|
ffgit_status | Branch, HEAD, an unfinished operation, staged, unstaged, untracked and conflicted files, linked worktrees |
ffgit_branches | Local and remote-tracking branches with upstream and ahead/behind counts |
ffgit_log | History on all branches or one ref, searchable by message, author, hash, path or code |
ffgit_show_commit | Author, date, full message, parents and changed files |
ffgit_diff | Working tree, staged, or between two commits, optionally for one path |
ffgit_blame, ffgit_line_history | Blame for up to 2,000 lines; git log -L for a line range |
ffgit_grep | Literal text in tracked files, in the working tree or at a commit |
ffgit_conflicts | The operation, which side is ours and theirs, and each conflict block |
ffgit_backups | The worktree’s recovery copies |
Files listed in .ffignore are reported apart, as expected changes.
Changes you approve
Section titled “Changes you approve”Agents can ask to stage, unstage, commit, create a branch, stash, resolve a conflict, continue an operation, and — destructively — skip a commit or abort an operation. Push, pull, discard and branch or worktree removal are not offered.
The running app carries out each change with a recovery copy, according to Changes from agents in Coding agents (MCP)…:
- Off — declines every change.
- Ask every time (default) — shows a prompt naming the agent, the repository and the change; it runs only after Approve.
- Allow without asking — runs changes at once, including repository hooks and filters.
Chat tools
Section titled “Chat tools”An orchestrator chat gets chat tools for its session: ffgit_chat_agents, ffgit_chat_start, ffgit_chat_send, ffgit_chat_wait, ffgit_chat_read, ffgit_chat_list, ffgit_chat_stop, ffgit_chat_answer and ffgit_chat_image. It starts Claude Code and Codex chats in your projects, messages them, waits on them and reads only the chats it started.
Agents outside ffgit get the same tools only by a grant: Coding agents (MCP)… → Outside agents that drive chats → Allow… picks the highest permission mode the agent may use (Up to Auto, Manual, Accept edits or Plan) and shows the claude mcp add … --chats <descriptor> command to copy. Take back ends the grant; its chats stay. On a Mac a script does the same:
ffgit-mcp chat start --chats <descriptor> --folder ~/code/app \ --agent codex --message "Run the tests and fix what fails" --waitWithout a grant the server offers no chat tools. Questions and approvals in the chats an outside agent starts stay yours.
Computer Use
Section titled “Computer Use”Open Coding agents (MCP)… → Computer use… and turn on Offer computer use to new chats for new Claude Code and Codex chats. macOS grants Accessibility to the ffgit Computer Use helper to read controls and Screen Recording to capture a window. ffgit asks separately for the chat and for each app. You can keep the answer for the chat or project, or choose Only this time.
A read returns one allowed window’s controls and text, with an optional screenshot. The agent sends these results to its provider with the chat. In Plan or Read only, the chat can inspect the window. In a mode that permits actions, it can also click, type, press keys or scroll. The floating Stop button cancels the action and, for a chat managed by ffgit, ends the current turn.
The helper’s own panel is blocked. The blocklist also covers known terminals, other agents’ apps, editors with terminals, System Settings, security prompts, Keychain Access and password managers. Allowing ffgit itself gives the agent access to its own approval cards and Privacy… settings.
For an outside agent, use Coding agents (MCP)… → Outside agents that use other apps. Choose Read windows only or Read and act. Each app still needs a separate decision. Window data goes to that outside agent. Computer Use permissions and revocation.