magi

mcp
Guvenlik Denetimi
Basarisiz
Health Uyari
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Low visibility — Only 8 GitHub stars
Code Basarisiz
  • rm -rf — Recursive force deletion command in cielago/scripts/checkout.sh
Permissions Gecti
  • Permissions — No dangerous permissions requested

Bu listing icin henuz AI raporu yok.

SUMMARY

Run coding agents across your machines: cielago moves and sends the work, magi shows it and suggests moves

README.md

magi

magi

See your coding agents on all your machines, and move their work to the one that fits.

magi · cielago · How it works · FAQ

In a Claude Code session on a Mac, the user asks for an agent on the machine that matches CI to run it and fix what fails. Claude reads the CI workflow, asks cielago for a Linux x86_64 machine and proposes the homelab, and the user agrees. A Claude starts on the homelab in its own copy of the repo, in a Rex tab named Pre-push CI. magi shows the Mac session's work running on the homelab. When the homelab Claude finishes, the Mac session is woken and reports one failing test, fixed, with CI passing and the fix on a branch to review.

Claude on a Mac sends CI to the Linux homelab. A Claude there fixes the failing test, and the Mac session is told when it's done.

This repo holds two apps:

  • cielago does what you ask. It hands a Claude Code or Codex session
    off to another machine, along with its conversation and uncommitted work.
    It also lets Claude send work there and be told when it's done. It works
    on its own, from any Claude Code session, over SSH.
  • magi is the layer on top. It's an overview of your agents and
    machines in one Rex pane: each
    machine's load, every session's state, and what needs you. When a machine
    is busy and another has room, it suggests moving a session there, and
    moves it through cielago when you say so.

magi

magi is an overview of your agents and machines, built to sit next to Rex's
sidebar. It handles any number of machines and needs no server or config
file.

  • It suggests moves. A machine counts as busy when its memory is 90%
    full, its CPU has averaged 80% for a minute, or its load reaches its core
    count. If a session there is using a lot of it and another machine has
    room, magi says so under that session. m moves it there through
    cielago, with its conversation and uncommitted work. magi never moves or
    starts anything by itself.
  • Every session, every machine. What needs you is at the top. Below
    that, each machine has a row with its load, and its sessions are listed
    under it, one line each, saying what each is doing.
  • Sent work stays in view. When Claude sends work with cielago, the
    session's line shows ⇢ homelab while it runs there, and the Claude
    doing it shows ← Pre-push CI, the session that sent it. With magi,
    cielago also knows each machine's live load when it picks one.
  • Answer from the overview. You can allow or deny a prompt, see a
    session's diff, or jump to its pane. Everything else happens in the
    session itself.
  • Limits and alerts. magi shows your 5-hour and weekly windows for Claude
    and Codex. It alerts you when a session needs you, a machine goes quiet or
    a plan nearly runs out.
  • Sessions start where there's room. A new session opens on whichever
    machine has room, in a Rex session named after its folder and a tab for
    its repo. Each subagent gets its own tab beside it, which shows its work
    live and closes when it's done.
  • Drive it from the Claude app. Sessions that magi starts run with Remote
    Control and are named after their repo and machine. That lets the Claude
    app on your desktop or phone talk to them. magi start prints the link.
  • Light and private. The agent idles at 0% CPU. Only devices signed in to
    your tailnet as you can connect.
curl -fsSL https://github.com/felitrejos/magi/releases/latest/download/install.sh | sh
magi agent install    # on every machine
magi                  # in a Rex pane

This installs magi and cielago. magi needs Tailscale
and Rex 0.1.0-preview.20 on every machine, and it only runs in Rex. While
Rex is in preview, magi follows only its latest release. To see Claude Code's
context use and plan windows, run magi statusline install where you use it.

cielago

cielago works on its own, over SSH, with or without magi. You ask, and it
moves a session to another machine or lets Claude send work there.

To move a session, say "hand this off to my homelab" in any Claude Code
session. When the turn ends, the conversation, branch, uncommitted changes
and listed secrets land there. The session carries on in a tab of its own.

A Claude Code session in Rex on a Mac is asked what its project does and explains a small Python library. Told to hand this off to the homelab, cielago moves it when the turn ends. The session opens in a tab on the homelab, and asked what it just said, it recaps its earlier answer.

A session handed off to the homelab still remembers what it said.

  • Hand off. cielago move <session> <host> sends a session, cielago pull <host> brings one back, and cielago doctor checks both ends.
    Codex sessions move the same way and reopen with codex resume.
  • Send work. Ask in plain words, as in the demo at the top. Claude can
    run a command on another machine, hand a task to a background Claude, or
    open a second session. Each runs in its own copy of your working tree, in
    a tab you can watch, and never touches your checkout. Rex shows a
    background Claude as Claude Code and a plain command as a process.
  • Claude asks first. It picks the machine by what the work needs: one
    that runs it natively, or else a Docker or Podman container on the least
    busy machine. Work that needs nothing in particular goes to the first
    machine in your list with room. It gets each machine's system,
    architecture and tools from magi, or over SSH for a machine magi doesn't
    report. Then it proposes the machine and waits for your yes, even when
    you tell it to pick.
  • Told when it's done. Claude runs cielago tasks wait <id> in the
    background, so the session is woken when the task finishes. A background
    Claude's changes come back as the branch cielago/task-<id>. cielago tasks lists what you've sent, cielago tasks show <id> prints a task's
    output, and cielago tasks cancel <id> stops it and removes its copy.
  • Ports follow. A project's .cielago.toml names what to start when it
    arrives and which ports to forward back, so localhost:3000 keeps working.
  • Rex or tmux. Sessions open in Rex when it's there and in tmux when it
    isn't. Both work over plain SSH.
curl -fsSL https://raw.githubusercontent.com/felitrejos/magi/main/install.sh | sh -s cielago
cielago init          # on each machine
claude plugin marketplace add felitrejos/magi && claude plugin install cielago@magi

From scripts and other tools

The same actions work from the command line. That way another orchestrator,
a script or Claude itself can use magi underneath:

magi sessions --json            # every session on every machine
magi send fix-login "now run the tests"   # queued until it's idle
magi answer fix-login yes       # or: magi choose fix-login 2
magi move fix-login homelab
magi start homelab ~/developer/stampr/stampr
magi events                     # what needs you, as JSON lines
magi subagents                  # every running subagent and the step it's on
magi recent fix-login           # its latest steps, one line each
magi wait fix-login             # returns once it's done or needs you

wait and recent are the cheap way for an agent to keep an eye on other
work. wait is one blocking command, so the agent doesn't have to check
again and again. recent gives a few lines per step instead of a whole
transcript. Both take a subagent too: magi wait fix-login "find auth".

magi mcp offers the same actions as MCP tools, so Claude can look after
your other sessions from Claude Desktop, Claude Code or any MCP client. In
Claude Code, run claude mcp add magi -- magi mcp. In Claude Desktop, add
this to claude_desktop_config.json:

{ "mcpServers": { "magi": { "command": "/Users/you/.local/bin/magi", "args": ["mcp"] } } }

The tools that only look (machines, sessions, subagents, recent work,
waiting, diffs, alerts) need no approval from the tools' side. The tools
that act (message, answer, choose, move, start) are marked as such, so your
client asks you first.

How it works

Each machine runs magi agent in the background (launchd or a systemd user
service). The viewer finds the agents through Tailscale and gets a small
sample from each one every second. The agent reads the files the coding
agents already write (~/.claude, ~/.codex) and follows Rex's events to
know each pane's state. When you answer a session, the agent on that machine
types your answer into the session's Rex pane.

Agent on an M4 MacBook CPU Memory
Nobody watching 0% 6 MB
32 viewers, every core busy 0.7% 7 MB
32 viewers, 5 live Claude sessions 2.9% 12 MB

FAQ

Can magi tell when Codex needs approval? Yes, from Codex 0.164 on.
Codex reports its prompts to Rex the way Claude Code does, and magi follows
them.

How do I move a Codex session? With cielago move, cielago pull or
m in magi, since Codex has no "hand this off". The conversation, branch,
changes and secrets move the same way. If Codex is in the middle of a turn,
it moves when the turn ends. If a plain codex is open in the same folder,
quit it first, because cielago can't tell which session it shows. The other
machine needs Codex installed, and cielago init asks for its path.

Does the status line slow Claude Code down? No. It prints first and
gives the agent at most 300 ms. It never replaces a status line you already
have.

What if Claude Code changes its files? They aren't a documented format.
magi notices, says so, and falls back to what the transcripts show.

What's next is in ROADMAP.md. What magi and cielago read, send
and change is in SECURITY.md. Licensed MIT.

Yorumlar (0)

Sonuc bulunamadi