Run Codex or Cursor instead of Claude Code

rev0 drives Claude Code by default. It can drive the OpenAI Codex CLI or the Cursor CLI instead, per ticket or for the whole board; this page shows how, and what each engine does differently.

Before you begin

Install the CLI on the machine the board runs on:

• Codex: npm install -g @openai/codex. It signs in with your ChatGPT or Codex subscription; the board wires up no API key.
• Cursor: run Cursor’s installer, curl https://cursor.com/install -fsS | bash, which puts the agent command on your PATH.
• Gemini: run the Antigravity installer, curl -fsSL https://antigravity.google/cli/install.sh | bash, which installs the agy command, then run agy once in a terminal and sign in with your Google account. The board cannot drive this sign in; Settings > Agents > Gemini auth only shows whether it is done.

Sign in from the board

1. Open Settings > Agents. Under Default model there is one panel per engine: Claude auth, Codex auth and Cursor auth.
2. In Codex auth (or Cursor auth), click Sign in.
3. Click Open the Codex authorize page and approve the sign in.
4. On the board’s own machine, the panel finishes by itself. From another device you land on a localhost page that cannot load: copy that whole address, paste it into Callback URL, and click Submit.

Result: the panel reads Signed in as your account, and any ticket that parked because the CLI was signed out goes back to To Do. The setup wizard’s AI engines screen has the same sign in. Signing in from a terminal (codex login, agent login) works too.

Route one ticket

1. On the Create ticket form, or in an open ticket’s toolbar, open Model.
2. Pick a GPT model to run on Codex, a model marked (Cursor) to run on Cursor, or one marked (Antigravity) to run on Gemini.

Result: the ticket runs on that engine. The engine follows the model: an id starting with gpt- or codex- runs on Codex, one starting with cursor/ runs on Cursor, one starting with agy/ runs on Gemini, and the rest run on Claude Code. No separate engine setting is needed.

Make it the board default

1. Open Settings > Agents.
2. In Default model, pick a GPT model or a (Cursor) model.
3. Click Save changes.

Result: every ticket with no model of its own runs on that engine.

To force the engine for every ticket whose model does not choose one, set REV0_ENGINE=codex or REV0_ENGINE=cursor. The prefixes that pick an engine, and each engine’s default model, live in config/engine_models.json.

Keep the CLIs up to date

Auto-update the agent CLIs, in Settings > Agents, is on by default. When the board starts, and every 6 hours after, it updates Claude Code, Codex, Cursor and Antigravity with each CLI’s own updater, and installs Cursor’s agent CLI if it is missing. When a CLI moves to a new version, the bell says so (for example Codex CLI auto-updated from v1.2.0 to v1.3.0.) with View changelog.

To turn it off, clear Auto-update the agent CLIs and click Save changes. The REV0_HARNESS_UPDATE variable overrides the setting (0 off, 1 on), and Mute notification kinds in Settings > Workspace silences the bell notice (An agent CLI was updated).

What is different on Codex

• The permission gate still applies. The board drives codex app-server, the Codex transport with an approval channel, so a Codex ticket gets the same permission cards, grants and containment as a Claude one. An implementing run may edit inside its worktree; anything that leaves it or touches the network escalates to the gate, and a plan run is read-only. REV0_CODEX_TRANSPORT=exec switches to the older transport, which cannot be gated, as an escape hatch.
• It reads its own MCP servers. A Codex run sees the board’s MCP server entries but not your Claude servers or account connectors; it reads ~/.codex/config.toml instead. See When a tool works in Claude Code but not for the agent.
• No plugin tools. A Codex run sees a trusted plugin’s hooks but not the tools it publishes.
• Its own thinking levels. Codex adds minimal and ultra to the shared levels; a level its model does not support is clamped down.

What is different on Cursor

Caution

A Cursor run skips the board’s approvals. The Cursor CLI’s print mode has no approval channel, so the board cannot hold a command for you: Cursor’s own sandbox and project permissions apply, but no permission card ever appears. Use Cursor for work you would let run unattended.

• The board’s MCP list does not reach it. Cursor finds MCP servers in its own configuration, so the board’s MCP server entries, including the board’s own tools, are not available to the run. The ticket’s timeline says so at the start of such a run.
• Plan runs are read only. A plan mode ticket runs Cursor in its ask mode, which cannot edit files; an implementing run can.
• No commands, skills or thinking levels. A Command ticket (a /skill) cannot run on Cursor, the CLI has no skills, and the ticket’s thinking level is ignored.

What is different on Gemini

The Gemini engine drives the Antigravity CLI, agy. Google’s own gemini CLI refuses a personal Google sign in, so it is not used.

Caution

A Gemini run skips the board’s approvals, and auto-approves every tool call. Antigravity’s print mode has no approval channel and prints nothing at all if a tool would have to ask, so the board runs it with all permissions granted. Use Gemini for work you would let run unattended.

• You only see the final answer. Print mode reports no tool calls, no thinking and no token counts, so the timeline shows one message per run and the ticket records no usage or cost. The diff still shows what changed.
• The board’s MCP list does not reach it. Antigravity reads ~/.gemini/config/mcp_config.json, so the board’s own tools are not available to the run. Settings > Connections > MCP servers can install an MCP server into that file, as it does for the other engines.
• Plan runs are read only by instruction. There is no read only mode; a plan run is told not to edit and runs with Antigravity’s terminal sandbox on.
• Every Antigravity model is listed. That includes the Claude and GPT-OSS models it serves; they run through agy, on your Google subscription.
• No commands, skills or thinking levels, as on Cursor. A model’s thinking level is part of its name, for example Gemini 3.1 Pro (High).

Each engine is compared side by side in What an agent sees and can do.

If something goes wrong

Cursor CLI is not logged in; sign in from Settings

The agent CLI is signed out. In Settings > Agents > Cursor auth, click Sign in. The Codex and Claude runs fail the same way when their CLI is signed out: sign in from their panel.

Cursor engine selected but the agent CLI is not installed

The ticket’s model runs on Cursor, but there is no agent command on the board’s PATH or in ~/.local/bin. Run Cursor’s installer, or pick a model for another engine.

Gemini engine selected but the agy CLI is not installed

The ticket’s model runs on Gemini, but there is no agy command on the board’s PATH or in ~/.local/bin. Run the Antigravity installer, or pick a model for another engine.

The login process stopped before it finished

The CLI’s sign in exited before you approved it, or printed no authorize link within 20 seconds. Open Login process output in the panel to read why, then click Sign in again.

Related

• Install rev0, for what the installer sets up and how to update the board.
• Connect an MCP client, for MCP servers on each engine.
• The permission gate, for what a gated engine checks.