M-x Machina is an Emacs package for running several coding agents side by side. Each one is named, filed in folders, color-coded by state, tied to its own Git worktree, and resumable from the same saved conversation after you restart Emacs.
An illustrative frame with fictional agents, not a screenshot. Click inside it, then use the same keys as the sidebar: one agent is waiting for approval, one has an unread reply, and one will finish while you watch.
Keyboard: j and k move, Enter opens or resumes, Tab shows details, right and left bracket jump between agents needing attention, z focuses, B opens the board, i diagnoses, question mark lists actions, x stops, u marks read, e opens eshell, q goes back. Shift+Tab or Escape leaves the frame.
Everything below exists in v0.1. Each idea starts from the key that does it in the sidebar.
Create a named agent in an existing checkout, or let Git create a new worktree and branch for it from the current commit. Uncommitted changes stay in the original checkout. Existing branches and directories are rejected, never reused.
Logical folders like Work/Billing Service organize the sidebar without moving any files. Collapsed folders still summarize working, waiting and unread agents inside them. Renaming an agent leaves its conversation ID alone.
Color means state. A purple * marks a reply you haven’t read, and it clears only after five continuous seconds on screen. ] and [ cycle through approval requests first, then unread replies.
The sidebar stays visible while you edit. B opens a board grouped by folder with Working, Waiting, Ready and Stopped columns. z fills the frame with one conversation, and pressing it again puts your windows back.
SQLite keeps each agent’s conversation ID, profile, worktree and branch. After a restart the dashboard loads only that metadata, and RET asks the backend for the same conversation. If that fails, nothing silently starts in its place, and no input is replayed.
Diagnostics checks the profile, executable, checkout, branch and saved identity without starting the agent or contacting the backend. W points a stopped agent at a moved worktree after you confirm. R retries once you’ve fixed the cause.
Creating an agent asks for agent, then account, then interface. All of them share the same sidebar, folders, worktree context and focus commands.
Native Emacs UI through ACP, reusing your agent-shell profiles and separate accounts. Tested with agent-shell 0.83.4.
Claude Code’s terminal UI, with per-run hooks that report status and session identity. Your Claude and project settings are never edited.
The same terminal integration through vterm’s native module, with history navigation and copy mode.
Press e for a separate shell in the agent’s worktree. It keeps its history and draft. It runs commands for you and is not an agent interface.
M-x mx-machina-messaging-mode starts a local message service. You, or an agent you instruct, use the mxm CLI to list peers and send a request. With --wait it returns that peer’s actual reply as JSON.
$ mxm list [{"id": "5d0c…", "name": "Work/Billing Service/Harness", "status": "live", "activity": "input", …}, …] $ mxm send 'Work/Billing Service/Harness' \ 'Do the new retry tests cover shutdown?' --wait Request ID: 2e6f… {"id": "2e6f…", "status": "completed", "response": "Yes. Two of the six cover shutdown mid-retry.", …}
list and every result are JSON. A timed-out wait exits with code 2 and leaves the request active.This is an experimental alpha. It’s worth knowing where the edges are before you rely on it.
Agents never assign each other work. Messaging carries only requests that someone explicitly sent.
Agents are processes owned by Emacs. They don’t keep running after Emacs exits.
The registry, sockets and message files are local to your machine. Credentials and history stay with each backend.
Resuming depends on the backend still having that history and on the original profile and account. When it fails, you see the failure instead of a replacement conversation.
Worktrees and branches are tracked, but pull requests and CI runs are not.
Integration tests cover all three interfaces offline. Real Claude messaging still needs an interactive pilot.
See the prioritized backlog for the path to dependable daily use.
Loading the package installs nothing globally. To try it without a model or account, start with the offline demo.
git clone https://github.com/eliraz-refael/m-x-machina.git ;; init.el (adjust the path) (add-to-list 'load-path "/path/to/m-x-machina/lisp") (require 'mx-machina) ;; then M-x mx-machina ; sidebar, press n for a new agent
M-: (sqlite-available-p)