Codex

@ask-llm/codex-mcp default gpt-5.6-sol → fallback gpt-5.6-terra

OpenAI's Codex CLI is the workhorse reviewer of the pair: strongest code reasoning for targeted reviews, architecture critique, and diff analysis. Claude asks Codex; Codex asks Claude back through @ask-llm/claude-mcp.

Best for: targeted code reasoning, architecture critique, and security review of specific files, the default workhorse reviewer. Not for: whole-repository reads beyond its context window (use Gemini or Antigravity), or fully offline/air-gapped use (it's a hosted model; use Ollama).

Installation

1. Provider CLI

npm install -g @openai/codex

2. Register the MCP server (Claude Code shown; see Quick Start for other clients)

claude mcp add --scope user codex -- npx -y @ask-llm/codex-mcp

Or install globally: npm install -g @ask-llm/codex-mcp

Prerequisites

1. Node.js v20.0.0 or higher
2. Codex CLI installed and authenticated

Tools

Tool	Purpose
ask-codex	Send prompts to Codex CLI. Optional reasoningEffort; omit sessionId for an ephemeral call, pass sessionId: "" to start a persisted thread, or pass a returned thread ID to resume it
ask-codex-edit	Propose structured code edits via a read-only codex exec --output-schema pass. Returns edit blocks for the calling client to apply
get-usage-stats	Per-session token totals + breakdowns. In-memory
ping	Fast connection test to verify MCP setup

ask-codex returns both human-readable text and a structured AskResponse (provider, response, model, sessionId, usage) via MCP outputSchema. Calls are ephemeral by default for privacy: a Thread ID from a call that omitted sessionId is not resumable. To opt into continuity, pass sessionId: "" on the first call; its returned Thread ID is the same value as structuredContent.sessionId, and can be passed back on later calls.

Models

When the default model hits quota, the executor retries once on the fallback and reports the actual model used in the response.

• Default: gpt-5.6-sol (GPT-5.6 Sol flagship)
• Quota fallback: gpt-5.6-terra, the balanced GPT-5.6 tier
• Overrides: ASK_CODEX_MODEL, ASK_CODEX_FALLBACK_MODEL, or the per-call model parameter

Key Features

• GPT-5.6 Sol access via the official Codex CLI
• Reasoning control: ordinary calls default to medium; /codex-review and /brainstorm use high; direct calls can request low, medium, high, xhigh, or max
• Native session continuity: omit sessionId for an ephemeral one-off call; pass sessionId: "" on turn one to persist a thread, then pass its returned thread_id on later turns. Follow-ups use codex exec resume <id> with the stable -c sandbox_mode="<mode>" grammar (zero replay cost, Codex retains state).
• Read-only, non-interactive sandbox: fresh calls use codex exec --sandbox read-only; resumed calls use the equivalent supported config override -c sandbox_mode="read-only". Both keep second-opinion, review, and edit-proposal calls from modifying the workspace. Codex exec is non-interactive by definition, so no approval prompt can hang the MCP subprocess. The optional sandbox: "workspace-write" parameter is a deliberate opt-out for flows that need Codex to write files (e.g. /codex-image); review flows must not set it.
• JSONL output parsing for structured responses + token usage
• Automatic quota fallback from GPT-5.6 Sol to Terra
• Structured AskResponse via outputSchema for programmatic clients
• Standard MCP transport works with 40+ clients

npm

• Package: @ask-llm/codex-mcp
• Binary: ask-codex-mcp