Second Opinion

An MCP server that lets AI coding agents consult other AI coding agents. Give Claude (or any MCP-compatible client) access to OpenAI Codex CLI, Opencode, and Claude Code as second-opinion tools.

What it does

Exposes a single MCP tool — get_second_opinion — that:

1. Sends a prompt to Codex, Opencode, or Claude
2. Waits for the agent to respond (with a configurable timeout)
3. Writes the response to a .md file
4. Returns the file path so the caller can read it

Note: No response body is returned inline. The caller must read the returned file path.

Execution model

This server shells out to external CLI tools. Each call spawns a subprocess that may read and write files within the specified working_directory.

Flags used per agent:

Agent	Flags
codex	--full-auto --ephemeral --skip-git-repo-check
opencode	--dangerously-skip-permissions
claude	-p --output-format text --dangerously-skip-permissions

These flags allow the agents to operate autonomously without interactive prompts. Only point them at directories you're comfortable with them accessing.

Quick start

1. Install runtimes and CLIs

Bun:

curl -fsSL https://bun.sh/install | bash

Codex CLI (requires Node 22+):

npm install -g @openai/codex

Opencode CLI:

curl -fsSL https://opencode.ai/install | bash

Claude Code CLI:

npm install -g @anthropic-ai/claude-code

2. Authenticate the CLIs

Codex uses your OPENAI_API_KEY environment variable:

export OPENAI_API_KEY=sk-...

Opencode has its own auth flow — run it once interactively to configure:

opencode

Claude Code uses your Anthropic API key or OAuth:

export ANTHROPIC_API_KEY=sk-ant-...

Or run claude auth to authenticate interactively.

3. Build

bun install
bun run build

This produces dist/second-opinion — a compiled binary. It still requires codex, opencode, and/or claude to be available on $PATH at runtime.

4. Configure your MCP client

Add to your MCP client config (e.g. Claude Code ~/.claude/settings.json):

{
  "mcpServers": {
    "second-opinion": {
      "command": "/path/to/dist/second-opinion"
    }
  }
}

Or run directly without building:

{
  "mcpServers": {
    "second-opinion": {
      "command": "bun",
      "args": ["run", "/path/to/src/index.ts"]
    }
  }
}

Tool reference

get_second_opinion

Parameter	Type	Default	Description
prompt	string	—	Question, code snippet, or task to send
agent	"codex" | "opencode" | "claude"	"codex"	Which agent to consult
working_directory	string	server's cwd	Directory the agent runs in. Agents may read/write files here. Must exist on disk.
timeout_seconds	number	120	Max seconds to wait (1–3600)
output_directory	string	system temp	Where to write the response file. Created automatically if omitted. If provided, must already exist.

Returns: The path to a .md file containing the agent's response.

On error: Returns isError: true. The response may still include a file path — check it, as partial output may have been written (e.g. on timeout).

Output file format

<!-- agent: codex | duration: 12.3s | exit: 0 | generated: 2026-01-01T00:00:00.000Z -->

[agent response here]

On timeout the status line will show TIMED OUT instead of an exit code. The body may contain partial output.

Example

Request:

{
  "prompt": "What's the most likely cause of an off-by-one error in a binary search?",
  "agent": "codex",
  "working_directory": "/my/project",
  "timeout_seconds": 60
}

Success response:

/tmp/second-opinion-codex-1234567890-ab12cd34.md

Read that file to see the agent's answer.

Troubleshooting

Symptom	Fix
Error spawning codex: ...	codex not on $PATH. Install it and verify with which codex.
Error spawning opencode: ...	opencode not on $PATH. Install it and verify with which opencode.
Error spawning claude: ...	claude not on $PATH. Install it and verify with which claude.
Agent exits with no output	CLI not authenticated. Run codex, opencode, or claude auth interactively once to check.
TIMED OUT in response file	Increase timeout_seconds or simplify the prompt.
Directory does not exist error	Supplied working_directory or output_directory doesn't exist on disk.

Development

bun run start            # Run without building
bunx biome check         # Lint + format check
bunx biome check --write # Auto-fix

Pre-commit hook runs Biome automatically via Husky + lint-staged.

License

MIT