Skip to main content
Glama

subcodex

MCP server for Claude Code to use OpenAI Codex as a subagent, with stall detection and auto-recovery.

中文文档

Why subcodex?

Claude Code and Codex have complementary strengths:

Role

Claude Code

Codex

Analogy

Product Manager who understands requirements

Technical expert in deep focus

Strengths

Planning, communication, context understanding

Coding, debugging, implementation

Best at

Breaking down tasks, writing specs, verification

Writing code, fixing bugs, refactoring

subcodex bridges these two, letting Claude Code orchestrate Codex as a subagent:

  • Claude Code handles the "what" and "why" (requirements, planning, verification)

  • Codex handles the "how" (implementation, debugging)

This combination delivers better results than either tool alone.

Related MCP server: codex-bridge

Features

  • Run Codex sessions with streaming progress

  • Automatic stall detection (configurable timeout)

  • Auto-recovery attempts when stalled

  • Progress logging to ~/.claude/codex-logs/

  • Thread continuation support via codex-reply

Usage Modes

Configure in your CLAUDE.md to control when subcodex is used:

Mode 1: Full Subagent

All code modifications go through subcodex. Claude only does analysis, planning, and verification.

## Subagent Mode: full-subagent

All code changes must go through `mcp__subcodex__run`:
- Claude: analyze, plan, write Codex Contract, verify results
- Subcodex: all file edits, code generation, refactoring

Mode 2: Directory-Based

Different directories are handled by different executors.

## Subagent Mode: directory-based

| Scope | Executor |
|-------|----------|
| `apps/web/**/*` | Claude direct |
| `apps/api/**/*`, `packages/**/*` | subcodex |
| docs, config | Claude direct |

Mode 3: Fallback

Claude handles everything, but falls back to subcodex on failure.

## Subagent Mode: fallback

- Claude attempts all code changes directly
- After 2 failed attempts → automatically use subcodex
- Windows file lock errors → immediately use subcodex

Installation

From npm

npx subcodex-mcp

From source

git clone https://github.com/G0d2i11a/subcodex.git
cd subcodex
pnpm install
pnpm build

Configuration

Add to your Claude Code MCP config (~/.claude.json):

{
  "mcpServers": {
    "subcodex": {
      "command": "npx",
      "args": ["-y", "subcodex-mcp"],
      "env": {},
      "type": "stdio"
    }
  }
}

Or for local development:

{
  "mcpServers": {
    "subcodex": {
      "command": "node",
      "args": ["/path/to/subcodex/dist/index.js"],
      "env": {},
      "type": "stdio"
    }
  }
}

Tools

run

Start a new Codex session.

Parameter

Type

Required

Description

prompt

string

Yes

The prompt to send to Codex

cwd

string

No

Working directory for the session

model

string

No

Model override (e.g., 'gpt-5.2')

sandboxMode

string

No

read-only, workspace-write, or danger-full-access

approvalPolicy

string

No

never, on-request, on-failure, or untrusted

level

string

No

Execution level: L1, L2, L3, L4 (for log naming)

stallTimeoutMinutes

number

No

Minutes of inactivity before detecting stall (default: 5)

maxRecoveryAttempts

number

No

Max auto-recovery attempts when stalled (default: 2)

reply

Continue an existing Codex conversation.

Parameter

Type

Required

Description

threadId

string

Yes

The thread ID from a previous session

prompt

string

Yes

The next prompt to continue the conversation

level

string

No

Execution level for log naming

stallTimeoutMinutes

number

No

Minutes of inactivity before detecting stall (default: 5)

maxRecoveryAttempts

number

No

Max auto-recovery attempts when stalled (default: 2)

Stall Detection

The server monitors Codex sessions for activity. If no events are received within the timeout period:

  1. Marks the session as stalled

  2. Attempts auto-recovery by sending a recovery prompt

  3. Retries up to maxRecoveryAttempts times

  4. Returns TIMEOUT status with needsUserInput: true if all recovery attempts fail

Handling needsUserInput

When the response contains needsUserInput: true, Claude should use AskUserQuestion to ask the user how to proceed. Add this rule to your CLAUDE.md:

## Subcodex Stall Handling

When `mcp__subcodex__run` returns `needsUserInput: true`:
- Use AskUserQuestion to ask the user how to proceed
- Options: retry, skip current task, manual intervention

Response Format

{
  "threadId": "abc123...",
  "level": "L2",
  "content": "Final response from Codex",
  "progressLog": "~/.claude/codex-logs/progress-L2-xxx-PASS.log",
  "stats": {
    "totalItems": 10,
    "commands": 3,
    "fileChanges": 2,
    "mcpCalls": 0,
    "usage": {
      "input_tokens": 1000,
      "output_tokens": 500
    }
  },
  "filesModified": ["create: src/foo.ts", "modify: src/bar.ts"],
  "recovery": {
    "attempted": false
  },
  "needsUserInput": false
}

When stalled and recovery fails:

{
  "threadId": "abc123...",
  "level": "L2",
  "content": "",
  "progressLog": "~/.claude/codex-logs/progress-L2-xxx-TIMEOUT.log",
  "recovery": {
    "attempted": true,
    "attempts": 2,
    "recovered": false,
    "lastError": "Still stalled after recovery attempt"
  },
  "needsUserInput": true
}

Result Levels

Log files are renamed with result level suffix:

  • PASS - Success (log file deleted)

  • FAIL - Command or file change failed

  • ERROR - Exception occurred

  • TIMEOUT - Stalled and recovery failed

Requirements

  • Node.js 18+

  • OpenAI Codex SDK credentials configured

License

MIT

Available Tools

2 tools
replyC

Continue a Codex conversation by providing the thread id and prompt

Inputs: prompt (required), threadId (required), level, maxRecoveryAttempts, +1 more.

runB

Run a Codex session with streaming progress, stall detection, and auto-recovery

Inputs: prompt (required), cwd, level, model, +4 more.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3.3/5.0

    Scored across 2 tools

    Disambiguation5/5

    The two tools are clearly distinct: 'run' starts a new Codex session, while 'reply' continues an existing one. The descriptions emphasize different purposes, so there is no ambiguity about which tool to use.

    Naming Consistency5/5

    Both tools are named with a single lowercase verb ('run' and 'reply'), following a consistent pattern. There is no mixing of naming conventions or unpredictable styles.

    Tool Count3/5

    With only 2 tools, the server feels thin, especially since it handles session management. However, for the narrow scope of starting and continuing a Codex conversation, the count is reasonable but borderline.

    Completeness3/5

    The core workflow of starting and continuing a session is covered, but there are notable gaps such as listing existing sessions, canceling a run, or retrieving status. This prevents full lifecycle coverage.

    Maintenance

    ActivityInactive
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers