amicus

Give your coding agent a second opinion from a different model.

amicus is one MCP server that fronts Codex, Kimi and Claude. The backend is a parameter, not a separate installation — so instead of running three servers and learning three tool surfaces, you ask for a consult, a review, or a delegated task and say which model should answer.

• Consult — a read-only second opinion on an approach, a bug, or a design.
• Review — a structured review of changes gathered from git, with findings tied to evidence.
• Delegate — a task implemented in a throwaway git worktree, returned as a diff you review. It is never applied to your working tree.
• Adversarial review — a fixed critic aimed at a plan, claim, or decision.

Long-running calls can be started in the background and polled, so a broad review does not block your session.

Install

Claude Code

/plugin marketplace add briandconnelly/amicus
/plugin install amicus@amicus

Codex

codex plugin marketplace add briandconnelly/amicus
codex plugin add amicus@amicus

Start a new session afterwards so the host loads the bundled skill and tools.

Any other MCP client

The plugins launch the server from a published release tag, pinned in .mcp.json. Point your client at the same command:

{
  "mcpServers": {
    "amicus": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/briandconnelly/amicus.git@v0.2.0", "amicus-mcp"]
    }
  }
}

Requirements

• Python 3.11 or newer, and uv.
• The CLI for each backend you enable, installed and authenticated: codex, kimi, claude. amicus drives those CLIs; it does not talk to the providers itself, and it stores no credentials.

Using it

Choose which backends are available with AMICUS_BACKENDS — a comma-separated list of codex, kimi and claude. Only the backends you name are enabled.

Then just ask, in your own words:

• "Get a second opinion on this approach from another model."
• "Have another model review my current changes."
• "Delegate this task to another model and show me the proposed diff."

In Claude Code there are also slash commands:

Start with /amicus:status — it tells you which backends are actually usable before you spend anything.

The tools

Nineteen tools, in five groups, one of them a deprecated alias. Everything paid has a free way to inspect it first.

Not every backend does everything:

A no is amicus's routing decision, not a claim about the model. delegate is absent on Claude by a deliberate review-only policy. adversarial_review is scoped to Claude for v1, but the verb is amicus's own: the prompt it builds and the result shape it returns are amicus's for any backend, and what Claude contributes is a critic stance layered on top. So nothing about Codex or Kimi stops either from reviewing adversarially — ask them through consult or review_changes and supply the adversarial stance yourself. What the verb adds is that stance fixed, plus its target and evidence carriers.

Asking a backend for something it does not support returns a feature_unsupported error naming the backend and the feature, rather than failing obscurely.

Safety

• A delegated diff is never applied. It runs in a throwaway git worktree and comes back for you to read and apply yourself.
• Every verb call spends the selected backend's quota. The dry runs and the discovery and job tools do not.
• Enabling a backend that can write makes the tool surface carry that backend's approval annotations — see docs/MIGRATION.md.
• Your prompts go to whichever provider you selected, through that provider's own CLI. Each backend discloses how it carries your text — visible on amicus_backends.

Configuration

AMICUS_BACKENDS is the one most people set. Beyond it, each backend takes an optional model, reasoning effort, and binary path (AMICUS_CODEX_MODEL, AMICUS_KIMI_REASONING_EFFORT, AMICUS_CLAUDE_BIN, and so on), and there are limits for timeouts, job retention and payload sizes. The full list is the env_vars array in .mcp.json.

Logging goes to stderr, and AMICUS_LOG_FILE mirrors amicus's own records to a file. The fastmcp and mcp libraries' records also reach stderr, but never below WARNING whatever AMICUS_LOG_LEVEL says, never the file, and with every unaudited FastMCP server record reduced to its level, logger and exception type. That is deliberate (ADR 0023): those libraries can log a caller's prompt text, so a FastMCP-side failure is diagnosed from the exception type and frames rather than its message.

Coming from codex-in-claude, moonbridge, or claude-in-codex

amicus replaces all three. docs/MIGRATION.md maps the old tool names onto the new ones.

Status and known limits

0.3.0 is the current release, and it is a breaking one: the discovery surface moved from amicus/0.1/schema-13 to amicus/0.1/schema-30 and stored job results from RESULT_FORMAT 4 to 6, so a job result 0.2.0 stored cannot be read after upgrading. docs/MIGRATION.md walks through what a caller written against 0.2.0 has to change, and CHANGELOG.md lists everything that moved. The server's wire contract is exercised by the test suite and by host captures against Claude Code and Codex, but:

• Two router-skill evaluation scenarios are unresolved — S6 (diff-safety wording) and S7 (approval friction) — recorded in docs/adr/0012-m6-packaging-decisions.md. Neither is a defect in the server's wire contract.
• A third-party backend distribution loads through the amicus.backends entry-point group, but cannot yet be enabled or called: AMICUS_BACKENDS and the backend parameter accept only the in-tree ids.

Development

uv sync
uv run prek install --prepare-hooks   # one-time local hooks

The gate is defined once, in AGENTS.md (Rules, item 2); CI runs it on every supported Python version.

License

MIT. See LICENSE.