Claude Code Host

@ask-llm/mcp is the canonical multi-host package and owns the Claude Code marketplace assets; @ask-llm/plugin remains a dependent bridge. This page covers the Claude Code adapter. For the Cursor-adapted pair skills (/codex-pair, /grok-pair) through Cursor Agent Skills/MCP, see Cursor Agent Host; for native Pi tools and lifecycle events, see Pi Host Support.

The Ask LLM plugin brings the second opinion into Claude Code itself: slash-command reviews (/codex-review, /multi-review), multi-model brainstorming (/brainstorm), and an opt-in continuous review hook (codex-pair) that checks every edit as you make it. Under the hood it adds review skills, brainstorm agents, and automated hooks.

Installation

Through setup (recommended)

Follow the Quick Start to install through setup. See Command compatibility for workflow installation and removal, and Skills for portable names and compatibility aliases. To install only the Claude Code plugin by hand:

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

From Source (development)

git clone https://github.com/Lykhoyda/ask-llm.git
cd ask-llm
yarn install && yarn build
claude --plugin-dir ./packages/llm-mcp

MCP Servers

The recommended cross-provider server is @ask-llm/mcp; follow the Quick Start to register it. Split provider packages (@ask-llm/codex-mcp, @ask-llm/grok-mcp, and the others) remain an advanced optimization for a richer per-provider tool surface.

The Claude Code plugin itself still ships only the Codex MCP registration used by review and pairing commands. Fully restart Claude Code after installation or upgrade, then run /mcp; plugin:ask-llm:codex should be connected and expose the exact ask-codex capability. Plugin agents and skills prefer that bundled ask-codex leaf, then a fully pinned unified ask-llm call (provider: "codex" plus model and Codex options), then the disclosed codex exec fallback.

/grok-pair relies on user-scoped servers instead of plugin bundling. Install the unified Ask LLM server (recommended: it exposes model-neutral ask-cursor-agent and the unified ask-llm tool, which pair skills call only with provider, harness, exact model, and effort pinned) and, optionally, the split Grok server for the ask-grok leaf:

ask-llm setup --host claude
claude mcp add --scope user grok -- npx -y @ask-llm/grok-mcp

Existing user-scoped Codex registrations remain supported if you prefer the shorter codex:ask-codex name. Register the other provider servers at user scope only when you want their split leaves:

claude mcp add --scope user antigravity -- npx -y @ask-llm/antigravity-mcp
claude mcp add --scope user ollama -- npx -y @ask-llm/ollama-mcp
claude mcp add --scope user gemini -- npx -y @ask-llm/gemini-mcp

If Codex registration is missing, provision the unified server first (ask-llm setup --host claude) or the split Codex leaf with claude mcp add --scope user codex -- npx -y @ask-llm/codex-mcp. If /mcp lists the server but it is disconnected, run ask-llm doctor and fully restart Claude Code. /sol-review preserves source-plugin and session-local MCP/settings context when reading the active claude mcp list inventory, treats failed health, an incomplete unified schema, or an MCP transport failure as unavailable, and discloses the explicit codex exec fallback.

What's Included

Skills (Slash Commands)

Command	Provider	Description
/multi-review	Antigravity + Codex	Parallel review with 4-phase validation pipeline and consensus highlighting
/gemini-review	Gemini	Get a second opinion on your current changes
/codex-review	Codex	Get a second opinion from GPT-6 Astra
/fable-review	Fable	Native isolated review, pinned to Fable
/sol-review	GPT-6 Sol	Model-pinned review through Codex
/grok-review	Grok	Explicit xAI API or Grok CLI review; metered/plan-aware, no harness or model fallback
/grok-pair	Grok	Consent-gated iterative reviewer through exact Cursor Agent, xAI API, or Grok CLI route; no fallback
/codex-pair	Codex	Continuous Claude hook dashboard; Cursor uses the separate on-demand persisted-session adapter
/ollama-review	Ollama	Local review, no data leaves your machine
/antigravity-review	Antigravity	Subscription-backed second opinion via Google agy (experimental)
/brainstorm	Explicit panel + Claude Opus evidence	Supports exact provider/harness/model routes, including a no-Gemini Grok + GPT-6 Sol panel through Cursor Agent; partial failures never become consensus
/brainstorm-all	All + Claude Opus	Requests all five external providers plus Claude Opus research; see the Antigravity execution gate
/compare	Multi (configurable)	Side-by-side raw responses from selected providers: no synthesis, no consensus extraction. Use when you want to see how each provider phrases the same answer

/codex-review and /sol-review require an installed, authenticated Codex CLI; the plugin supplies their MCP registration. /grok-review, /ollama-review, /antigravity-review, and bare-provider /brainstorm routes require their respective configured tools and credentials. Routed /brainstorm ...@cursor-agent:<exact-id> participants use the packaged model-neutral Cursor runner and require an authenticated Cursor Agent CLI; they do not silently use provider MCP tools as fallback.

Looking for continuous background review (not a slash command)? See codex-pair, a PostToolUse hook that runs Codex against every file edit when a project has opted in via a marker file. It's the recall-first complement to /codex-review.

Agents

Agent	Description
gemini-reviewer	Isolated Gemini code review with confidence-based filtering
codex-reviewer	Isolated Codex code review with confidence-based filtering
fable-reviewer	Native read-only Fable review with source validation
sol-reviewer	GPT-6 Sol review through Codex with source validation
grok-reviewer	Grok review through the selected API/CLI harness with source validation and no fallback
ollama-reviewer	Local Ollama code review, no data leaves your machine
antigravity-reviewer	Subscription-backed Antigravity (agy) code review, experimental
brainstorm-coordinator	Researches before dispatch, then synthesizes explicit participants. In exact Grok + Sol mode Claude is a non-voting verifier and two-model consensus requires both requested models to succeed.

Hooks

Hook	Trigger	Action
codex-pair PostToolUse	After every Edit/Write/MultiEdit	Opt-in. Self-gates on .codex-pair/context.md marker file. Zero cost without the marker. With marker: edits are debounced into a settle window, a detached worker reviews the settled file state, and HIGH/MED verdicts surface to Claude on a later edit, the next user prompt, or at turn end. See Codex Pair for opt-in steps and cost characteristics
codex-pair-prompt-drain UserPromptSubmit	On every user prompt	Drains queued codex-pair verdicts that finished mid-turn so they reach Claude without waiting for the next edit
codex-pair-stop-gate Stop	At turn end	Drains remaining queued verdicts (no opt-in needed). With blockOn: HIGH in the marker frontmatter (opt-in, default OFF), blocks turn-end while unaddressed HIGH findings or in-flight reviews remain
codex-pair-session SessionStart / SessionEnd	At Claude session boundary	Manages pause, debounce, and broker state; see Codex Pair

The hook has zero workspace imports, required so it runs from marketplace git-subdir installs that don't run npm install. A previous PreToolUse Gemini-review pre-commit hook was removed because continuous codex-pair review covers the same need with higher recall; use /gemini-review or /codex-review on demand for explicit pre-commit review instead.

CLI Binaries (source builds only)

These commands are available after cloning and building the plugin locally. Marketplace git-subdir installs do not build or ship the generated dist/ binaries.

Command	Description
ask-gemini-run	Pipe code or prompts directly to Gemini CLI
ask-codex-run	Pipe code or prompts directly to Codex CLI
ask-grok-run	Pipe code or prompts to the explicitly configured Grok API/CLI harness
ask-brainstorm-run	Run the exact Grok + GPT-6 Sol panel with repeated provider@harness:exact-model-id participant specs and structured partial-failure output
ask-ollama-run	Pipe code or prompts directly to local Ollama

How It Works

The plugin uses several Claude Code integration points:

1. plugin.json + .mcp.json: Explicitly declares the Codex MCP component for Claude Code plugin sessions (see Installation); Cursor/Grok routes use user-scoped @ask-llm/mcp and @ask-llm/grok-mcp registrations, and the separate .cursor-plugin/plugin.json + unified-only mcp.json adapter uses Cursor's supported surfaces
2. Skills (skills/): User-invocable slash commands that trigger review or brainstorm workflows
3. Agents (agents/): Handle the actual interaction with each provider using confidence-based filtering (80%+ threshold). Agents read CLAUDE.md for project conventions when available.
4. Hooks (hooks/): Run the opt-in codex-pair continuous review pipeline: per-edit PostToolUse reviews, verdict drains on user prompts and at turn end, the opt-in Stop gate, and session lifecycle
5. Source-build CLI binaries (src/): After a local build/link, enable piped analysis from shell: git diff | ask-gemini-run "review this"

Requirements

• Claude Code installed and authenticated
• Codex CLI authenticated, required for /codex-review and brainstorm with Codex
• Ollama running locally, required for /ollama-review and brainstorm with Ollama
• Gemini CLI authenticated (gemini login), required for Gemini features
• For /brainstorm, at least two providers should be available for meaningful synthesis

Source

• Pi: ask-llm setup --host pi (guide)
• Marketplace: /plugin marketplace add Lykhoyda/ask-llm then /plugin install ask-llm@ask-llm-plugins
• Source: packages/llm-mcp