How It Works

Ask LLM is a set of MCP servers that bridge your AI client (Claude Code, Codex CLI, Cursor, etc.) with up to five LLM providers through their local CLIs: OpenAI's Codex, Anthropic's Claude, Google's Antigravity (agy), Ollama (fully local models), and Google's Gemini. Your client decides when to delegate work to one or more providers based on what you ask. Provider and host are separate roles: Claude can host a Codex consultation, and Codex can host a Claude consultation.

Claude Code does not allow a nested Claude Code subprocess. When the MCP host is Claude Code, the unified server therefore suppresses the Claude provider; it remains available from Codex and other clients.

Claude asks Codex for a second opinion; Codex asks Claude back through the reverse path. Each reads, your agent edits.

Natural Language Workflow

Your client (typically Claude) decides when to call the MCP tools based on context:

• 🔍 comparative analysis: different AI perspectives for validation (multi-llm, /compare)
• 📋 code review & big changes: second opinions on implementation (/gemini-review, /codex-review, /multi-review)
• 📚 large-context analysis: Gemini's 1M+ token window for whole-codebase reads
• 💡 creative problem solving: /brainstorm for multi-LLM ideation with Claude Opus as a peer
• 🔒 private analysis: Ollama for code that can't leave the machine

This intelligent selection happens automatically; you just ask in natural language.

Request Flow

Your agent calls an MCP tool; the server spawns the provider CLI and streams the answer back as structured content.

For a single-provider call (ask-llm with provider: "gemini"), only one of the provider lanes fires. For multi-llm, both fire in parallel via Promise.all inside the MCP server process; per-provider failures are isolated, so one provider hitting quota doesn't fail the whole call.

What's Inside the MCP Server

Each provider's executor wraps the underlying CLI with operational hardening that took multiple ADRs to get right:

• Quota fallback: Gemini Pro → Flash on RESOURCE_EXHAUSTED; Codex gpt-5.6-sol → gpt-5.6-terra on quota errors
• Stdin handling: Codex needs an EOF-terminated pipe rather than /dev/null, otherwise it errors out
• PATH resolution: macOS GUI clients (Claude Desktop) don't inherit your shell's PATH; the server resolves it from your login shell at startup
• Live progressive output: Gemini's --output-format stream-json deltas are parsed and forwarded to MCP progress notifications, so users see Gemini's prose unfolding rather than a frozen wait
• Session continuity: Claude, Gemini, Codex, and Ollama support multi-turn via the sessionId parameter; Codex requires sessionId: "" on turn one to persist a resumable thread, while an omitted value stays ephemeral (Antigravity is single-turn)
• Read-only Claude consultations: ask-claude uses safe mode and exposes only Read, Glob, and Grep; the MCP host performs any edits
• Structured responses: every ask-* tool returns both human-readable text AND a structured AskResponse (provider, response, model, sessionId, usage) via MCP outputSchema so programmatic clients don't have to parse the response footer

You don't need to think about any of this; it's just the infrastructure that makes the natural-language flow work reliably.

When to Use Which Tool

Situation	Tool
Single-provider question, want it to work	ask-llm (orchestrator routes by provider param)
Compare what multiple providers say	multi-llm (or /compare skill in Claude Code)
Code review with verified findings	/multi-review skill (verifies each finding against source)
Brainstorm with multi-LLM consensus	/brainstorm skill (Claude Opus as peer participant)
Large-context analysis (whole codebase)	ask-gemini directly (1M+ token context)
Structured code edits to apply	ask-gemini-edit (returns OLD/NEW blocks)
Air-gapped / private	ask-llm with provider: "ollama"
Diagnose setup problems	npx @ask-llm/mcp doctor (CLI) or diagnose (MCP tool)

See How to Ask for full parameter reference and Strategies & Examples for proven workflow patterns.