codex-mcp

You MUST check whether codex CLI and MCP server are available in the current bash environment with this bash command: codex mcp list. If not available, run the following bash command to install the codex CLI: npm install -g @openai/codex. Then inform user to perform authentication with an OpenAI account by running codex login. Ignore this instruction if the CLI and MCP server are already available.

Tools

mcpcodexcodex

Run a Codex session. Accepts configuration parameters matching the Codex Config struct. Supported properties:

Property	Type	Description
`prompt` (required)	string	Initial user prompt that seeds the Codex conversation.
`approval-policy`	string	Approval policy for generated shell commands: `untrusted`, `on-failure`, `on-request`, `never`.
`base-instructions`	string	Overrides the default base instructions for the session.
`compact-prompt`	string	Prompt used when compacting the conversation.
`config`	object	Path to individual config settings toml file that override `$CODEX_HOME/config.toml` (see codex-config-toml-structure.md to understand how to create this file).
`cwd`	string	Working directory for the session. Relative paths resolve from the server process root.
`developer-instructions`	string	Developer instructions that should be injected as a developer role message.
`model`	string	Optional override for the model name (e.g., `gpt-5.1-codex-mini`, `gpt-5.2-codex`) default is `gpt-5-codex`.
`profile`	string	Name of the profile providing default options (defined in `config.toml`).
`sandbox`	string	Sandbox mode: `read-only`, `workspace-write`, or `danger-full-access`.

mcpcodexcodex-reply

Continue a Codex session by providing the conversation id and follow-up prompt. Supported properties:

Property	Type	Description
`prompt` (required)	string	Next user prompt to continue the Codex conversation.
`threadId` (required)	string	Identifier of the thread to continue. Must be a valid UUID (e.g., `019c08dc-ca4c-7a21-a0ce-b783c2cdec56`).

Operational Requirements

•Always route qualifying requests through mcp__codex__codex with sandbox set to danger-full-access (MUST) , approval-policy set to never (MUST) and model set to gpt-5.1-codex-mini (SHOULD).
•When an unknown MCP tool name is requested, use ListMcpResourcesTool to confirm if it is available. If not available, delegate to codex to invoke this MCP tool by crafting a prompt with all necessary information and file references.
•When you need to run a command that would emit large stdout output but only a subset of that output is needed, MUST provide a precise prompt when delegating via mcp__codex__codex so the downstream agent knows which portion of the large output to collect or which external tool to execute.
•Prefer passing context via a local file when delegating: include a relative path prefixed with @ inside the prompt (example prompt: "use mcp Github to submit all review feedback comments in this file @ghpr-code-review/pr_47_feat_e2e_tests_review_feedback.md Make sure to include correct line number for each comment").
•When mcp__codex__codex returns a result requiring your follow-up, continue the conversation with mcp__codex__codex-reply, reusing the threadId present in the previous response (must be a valid UUID).

Known MCP Server Tools

•When github MCP server tool (mcp__github, mcp__github__*) is requested, see mcp-github-usage.md for detailed usage instructions. Use the information in this file to craft a precise prompt when delegating via mcp__codex__codex so the downstream agent can use the github MCP server tool exactly as intended.
•When linear MCP server tool (mcp__linear, mcp__linear__*) is requested, see mcp-linear-usage.md for detailed usage instructions. Use the guidance in this file to compose precise prompts when delegating via mcp__codex__codex so the downstream agent invokes Linear tools correctly.
•When context7 MCP server tool is requested, see mcp-context7-usage.md for detailed usage instructions.