# OpenAI Codex


# Set up banish for OpenAI Codex

banish runs as an MCP server. Register it in the Codex CLI and every command it
runs through banish comes back compacted, so Codex spends far fewer tokens on
verbose shell and tool output.

## Prerequisites

- banish installed. See [install](/docs/getting-started/install) and confirm with
  `banish version`.
- OpenAI Codex CLI, which supports MCP servers over stdio.

## Register banish

Codex uses TOML, not JSON. Add banish with the CLI:

```sh
codex mcp add banish -- banish serve
```

Or add it by hand to `~/.codex/config.toml`:

```toml
[mcp_servers.banish]
command = "banish"
args = ["serve"]
```

The table key is `mcp_servers` (TOML, snake case) - do not paste the JSON form used
by other agents. If Codex cannot find banish on your PATH, use the absolute path
from `which banish` as the `command`.

## Verify

Start Codex and list its tools. banish exposes every verb in your extensions as an
MCP tool - with the defaults you get 45-plus tools such as `banish_gs`,
`banish_dps`, and `banish_kpods`. If they appear, banish is wired in.

## What you get

Codex now reads compacted output through banish. See
[MCP server](/docs/concepts/mcp-server) for how verbs become tools, the
[benchmarks](/docs/reference/benchmarks) for measured savings, and
[set up an agent](/docs/getting-started/set-up-an-agent) for the general flow.

Other agents: [Gemini CLI](/docs/integrations/gemini-cli),
[Windsurf](/docs/integrations/windsurf), [Cline](/docs/integrations/cline),
[GitHub Copilot](/docs/integrations/copilot).