Ask LLM

MCP servers + Claude Code/Pi host package for AI-to-AI collaboration

Get a second opinion before you ship. Ask LLM lets your AI assistant — Claude Code, Codex CLI, Cursor, Claude Desktop, or any of 40+ MCP clients — consult a second model to review your code, debate a plan, or catch a bug it might have missed. Pick the reviewer that fits: OpenAI Codex (GPT-5.6 Sol → Terra), Anthropic Claude (Opus → Sonnet), xAI Grok 4.6 via API or Grok CLI (no fallback), Google Antigravity (agy), a local Ollama model, or Gemini (1M+ token context). Standard MCP, no prompt hacks.

⚠️ Gemini CLI goes enterprise-only on 2026-06-18: From that date Google restricts Gemini CLI to Gemini Code Assist Standard/Enterprise seats, and free, Google AI Pro, and Ultra accounts lose access. @ask-llm/gemini-mcp still installs, but a non-enterprise account then surfaces actionable guidance instead of output. Free/Pro users: switch to ask-antigravity (the Google-sanctioned successor, subscription-backed via Google AI Pro/Ultra), ask-codex, ask-claude, or ask-ollama. Announcement

Why a second opinion?

Your primary AI is confident — but confidence isn't correctness. A second model, with no stake in the first one's answer, catches what it missed.

• Second opinion on code — before you commit to an approach, have another model review it independently.

• Debate a plan — send an architecture proposal for critique, alternatives, and trade-off analysis.

• Review a diff — have a different model analyze your changes to surface issues your primary AI glossed over.

• Read more than fits — Gemini and Antigravity's large context windows ingest whole codebases at once.

• Keep it local — run reviews through Ollama when nothing can leave your machine.

Related MCP server: SystemPrompt Coding Agent

In action

You:    ask codex to review src/auth.ts for security issues
Codex:  ⚠ verifyToken() compares tokens with === — not timing-safe (line 42)
        ⚠ the session cookie is missing a SameSite attribute
Claude: Good catches — applying both fixes to src/auth.ts.

One prompt. A second model reviews independently; your assistant applies the fix — no copy-paste between tools.

Quick Start

Claude Code

# All-in-one — auto-detects installed providers
claude mcp add --scope user ask-llm -- npx -y @ask-llm/mcp

claude mcp add --scope user gemini -- npx -y @ask-llm/gemini-mcp
claude mcp add --scope user codex -- npx -y @ask-llm/codex-mcp
claude mcp add --scope user grok -e XAI_API_KEY="$XAI_API_KEY" -- npx -y @ask-llm/grok-mcp
claude mcp add --scope user ollama -- npx -y @ask-llm/ollama-mcp
claude mcp add --scope user antigravity -- npx -y @ask-llm/antigravity-mcp

Pi

Pi is a host harness, not another consulted provider. Install the canonical dual-host package; it exposes the shared skills plus native provider tools because Pi intentionally has no built-in MCP client:

pi install npm:@ask-llm/plugin
pi list

Invoke /skill:codex-review, /skill:multi-review, /skill:compare, /skill:brainstorm, or describe the workflow naturally. The package registers ask-codex, ask-gemini, ask-grok, ask-ollama, ask-antigravity, model-neutral ask-cursor-agent, and deterministic concurrent ask-multi tools. Provider CLI authentication is unchanged and separate from Pi's host-model login.

Pi codex-pair requires the repository marker, Pi project trust, and interactive user-owned consent via /codex-pair; a committed marker alone never authorizes source transfer or cost. Pi surfaces findings non-blockingly and does not claim Claude's blocking Stop-gate or one-shot print parity. Independent fable-review remains Claude Code-only and is excluded from Pi.

Update/remove with pi update npm:@ask-llm/plugin and pi remove npm:@ask-llm/plugin. See the Pi host guide for temporary/project-local installs, trust, data transfer, troubleshooting, and the full compatibility matrix.

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "ask-llm": {
      "command": "npx",
      "args": ["-y", "@ask-llm/mcp"]
    }
  }
}

{
  "mcpServers": {
    "gemini": {
      "command": "npx",
      "args": ["-y", "@ask-llm/gemini-mcp"]
    },
    "codex": {
      "command": "npx",
      "args": ["-y", "@ask-llm/codex-mcp"]
    },
    "ollama": {
      "command": "npx",
      "args": ["-y", "@ask-llm/ollama-mcp"]
    }
  }
}

Cursor (.cursor/mcp.json):

{
  "mcpServers": {
    "ask-llm": { "command": "npx", "args": ["-y", "@ask-llm/mcp"] }
  }
}

Codex CLI (~/.codex/config.toml):

[mcp_servers.ask-llm]
command = "npx"
args = ["-y", "@ask-llm/mcp"]

For the focused Codex → Claude second-opinion path:

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

Any MCP Client (STDIO transport):

{ "command": "npx", "args": ["-y", "@ask-llm/mcp"] }

Replace @ask-llm/mcp with @ask-llm/codex-mcp, @ask-llm/claude-mcp, @ask-llm/grok-mcp, @ask-llm/antigravity-mcp, @ask-llm/ollama-mcp, or @ask-llm/gemini-mcp for a single provider.

Migrating from the old package names

All public MCP packages now live in the @ask-llm npm organization. The old package names are deprecated, but their executable names are unchanged. Update the package argument in your MCP config; commands such as ask-codex-mcp and ask-llm-mcp doctor keep working after a global install.

See the installation guide for the complete package-to-executable mapping.

Choose your reviewer

Host Package: Claude Code and Pi

@ask-llm/plugin is one package, version, release lifecycle, and canonical skill corpus. Claude Code loads its existing marketplace agents/hooks; Pi loads explicit native tools, portable skill adapters, and a thin lifecycle extension.

Claude Code Plugin

The Ask LLM plugin adds multi-provider code review, brainstorming, and automated hooks directly into Claude Code:

/plugin marketplace add Lykhoyda/ask-llm
/plugin install ask-llm@ask-llm-plugins

What You Get

The review agents use a 4-phase pipeline inspired by Anthropic's code-review plugin: context gathering, prompt construction with explicit false-positive exclusions, synthesis, and source-level validation of each finding.

See the plugin docs for details.

Prerequisites

• Node.js v20.0.0 or higher (LTS)

• At least one provider:

  • Codex CLI — installed and authenticated

  • Claude Code CLI — installed and authenticated (for Codex/other clients consulting Claude)

  • xAI API — set XAI_API_KEY for the default metered Grok harness; or install/authenticate official Grok Build and set ASK_GROK_HARNESS=grok-cli

  • Cursor CLI — optional model-neutral harness; authenticate and choose an exact ID from agent --list-models

  • Antigravity CLI (agy) >=1.1.5 — installed and logged in once (Google AI Pro/Ultra); verify with agy --version

  • Ollama — running locally with a model pulled (ollama pull qwen3.6:27b)

  • Gemini CLI — npm install -g @google/gemini-cli && gemini login (enterprise-gated from 2026-06-18)

MCP Tools

Session-capable ask-* tools accept an optional sessionId parameter and return a structured AskResponse (provider, response, model, sessionId, usage) via MCP outputSchema alongside the human-readable text. Codex requires sessionId: "" on the first call for a resumable thread; omitting it makes that call ephemeral. The orchestrator (@ask-llm/mcp) also exposes usage://current-session as an MCP Resource for live JSON snapshots.

Usage Examples

ask codex to review the changes in src/auth.ts for security issues
ask claude for an independent opinion on this architecture (from Codex or another MCP client)
ask antigravity to debate this architecture plan in docs/design.md
ask ollama to explain src/config.ts (runs locally, no data sent anywhere)
ask gemini to summarize @. the current directory (1M+ context, @ is Gemini-only)
use multi-llm to compare what codex and gemini think about this approach

CLI Subcommands

The orchestrator binary (@ask-llm/mcp) supports two CLI modes alongside the default MCP server:

# Interactive multi-provider REPL — switch providers, persist sessions, see usage live
npx @ask-llm/mcp repl

# Diagnose your setup — Node version, PATH, provider CLI versions, env vars
npx @ask-llm/mcp doctor                       # human-readable
npx @ask-llm/mcp doctor --json                # established full JSON, exit 1 on error
npx @ask-llm/mcp doctor --format toon         # bounded, versioned agent-facing TOON pilot
npx @ask-llm/mcp doctor --format toon --full  # full TOON escape hatch

The REPL ships sessions per provider (/provider gemini, /provider codex, /new, /sessions, /usage) and inherits all the executor behavior (quota fallback, stream-json output for Gemini, native session resume).

Models

Gemini, Codex, Claude, and Antigravity automatically fall back under their documented provider-specific conditions. Grok and Ollama never substitute a model: Grok preserves the explicit API/CLI harness catalog selection, while Ollama preserves the locally pulled model — if the requested model isn't pulled, it returns a clear ollama pull error.

Documentation

• Docs site: lykhoyda.github.io/ask-llm

• AI-readable: llms.txt | llms-full.txt

Contributing

Contributions are welcome! See open issues for things to work on.

License

MIT License. See LICENSE for details.

Disclaimer: This is an unofficial, third-party tool and is not affiliated with, endorsed, or sponsored by Anthropic, Google, OpenAI, or xAI.