Claude Code and Codex in Emacs.

When working with Claude Code, you often want multiple agents working simultaneously:

• One agent refactoring the auth module
• Another writing tests
• A third updating documentation

But this creates problems:

File conflicts

Agents might edit the same files.

Context sharing

How do agents know what others are doing?

Permission chaos

Multiple agents ask for input at once.

Independent review

How do you get one model to review another model's work?

Magnus solves all of this.

Magnus does not replace Claude Code or Codex with a generic chat buffer. Each agent runs in its provider's full TUI inside vterm. Streaming output, commands, approvals, and the terminal composer continue to work normally.

1. 1

   Create a named agent

   Choose Claude Code or Codex, then choose a Git project. Magnus gives the session a memorable name such as swift-fox.

   c Claude Code   ? X Codex

2. 2

   Work in the real terminal interface

   Visit an agent whenever you want to inspect, redirect, or continue its work. Magnus remains orchestration around the provider—not another provider UI.

   RET visit   m send a message

3. 3

   Return to the project view

   See active work, health, attention, shared discoveries, and decisions. Press a to visit the next agent waiting for you.

4. 4

   Ask another model to review the commits

   When a unit of work is committed, Magnus can run a fresh reviewer and display anchored findings beside the exact diff.

   v RET request a review

Named sessions

Create, visit, steer, rename, archive, and resurrect agents. Captured provider sessions return with their conversation intact when the CLI supports it.

Advisory coordination

Agents record ownership, discoveries, decisions, and @name messages in a shared project journal. They are instructed to follow the protocol; Magnus does not pretend file conflicts are impossible.

Attention and health

Permission and input requests enter one queue. Health indicators show which terminals are active, quiet, stuck, or gone.

Context and traces

Keep shared source material in a project buffer and follow the provider history Magnus can record, including visible engineering journals and messages.

Independent reviews

Send exact committed evidence to a headless reviewer from the other provider, read findings in a folding Magit-style diff, and continue with the same reviewer across rounds.

Magnus requires Emacs 28.1+, vterm, transient, and magit-section. Install and authenticate Claude Code, Codex, or both before launching them through Magnus.

MELPA

(use-package magnus
  :ensure t
  :bind (("C-c m" . magnus)
         ("C-c M" . magnus-create-instance)))

Or run M-x package-install RET magnus RET.

Latest revision from GitHub (Emacs 29+)

M-x package-vc-install RET https://github.com/hrishikeshs/magnus RET

First run

1. Run M-x magnus-doctor to check the local setup.
2. Run M-x magnus.
3. Press c for Claude Code, or ? then X for Codex.

The getting-started guide covers provider setup, straight.el, quelpa, upgrades, and troubleshooting.

Claude Code and Codex are built in. To use another CLI, write a provider adapter for the lifecycle, messaging, trace, and headless operations it supports. Magnus's shared layers continue to provide identity, persistence, coordination, attention, and status.

Start with the small dispatch interface in magnus-provider.el and the complete Codex implementation in magnus-provider-codex.el.

Magnus launches the configured provider CLIs and uses their normal authentication and network behavior. There is no Magnus daemon or replacement chat runtime.

Claude Code ─┐
             ├─ Magnus / Emacs ─ status, identity, coordination
Codex ───────┘                  attention, context, traces, review

See the architecture guide for module ownership, lifecycle boundaries, storage, and provider integration.