Start with Codex

Mjolnir is a self-hosted power frontend for Codex and Claude. Codex remains the coding agent that plans, edits, runs tools, and answers; Mjolnir supplies the interface and operating environment around that session. Setting up Claude instead? See Start with Claude.

What Mjolnir adds

Capability	What it changes
First-class coding teams	Run Codex, Claude, Codex coder + Claude reviewer, or Claude coder + Codex reviewer without changing tools or workflows
Integrated adversarial review	Hold a workspace-changing turn while an independent reviewer challenges the diff before the turn is released
Self-hosted remote control	Drive the session from another browser or device while the repository and Mjolnir server remain on your machine
Worktree-first workflow	Start Codex in a linked Git worktree so agent changes stay separate from your current checkout and easy to inspect
Cross-platform voice	Add locally transcribed prompts with Ctrl-R on supported desktop platforms

Mjolnir, its transcript storage, remote server, and Mjolnir-hosted tools run on infrastructure you control. Codex model requests still go to OpenAI under the terms and data boundaries of your Codex account. Mjolnir does not make the model service itself self-hosted.

Recommended setup

Install and authenticate the official Codex CLI first:

npm install -g @openai/codex
codex login

Then install Mjolnir, open a repository, and run:

mj

First launch opens Mjolnir’s onboarding on the Team tab. Choose Codex to keep primary, subagent, and review model selection on Auto while constraining all three seats to the Codex adapter. Then confirm:

1. OpenAI / ChatGPT reports that you are signed in.
2. The Codex ACP server is detected and enabled.
3. The primary model resolves to a Codex model. Keeping the model on Auto lets Mjolnir choose among currently launchable ranked Codex models.
4. Discrete review is on if you want workspace-changing turns reviewed before they are released.

Press Ctrl+Tab during a session to switch among all four teams, or return to the same choice on the Team tab in /mjconfig. Start a new session after switching so the new coder and reviewer routes apply together.

Use /agents after launch to record the model and adapter that actually bound to each seat. The current session keeps those choices until /new or /clear.

A safe first session

For change-producing work, start in an isolated linked worktree:

mj --worktree

Try a small request with an observable validation step. Mjolnir will show Codex’s messages, thoughts, tools, permissions, and any subagent or review activity in one transcript. Run the checked 10-minute evaluation for an end-to-end exercise in a disposable repository.

Go beyond the terminal

• Press Ctrl-R to dictate a prompt on a supported desktop installation.
• Run mj server to open the same session through the self-hosted remote viewer.
• Ask Codex to launch a subagent for bounded parallel work.
• Run /review recent for a findings-only review of the latest change-producing turn.
• Run mj --print for scripts and machine-readable output.

Continue with Remote control, Voice dictation, or Delegation and adversarial review.

Pair Codex with Claude

The mixed teams are where Mjolnir earns its keep: Codex coder + Claude reviewer keeps Codex in charge of the turn while Claude challenges every changed turn from an independent session, and the reverse team swaps the roles. See Teams and adversarial review, or Other agents and models for model resolution details.