Ask LLM

How to Ask ​

You don't need to memorize commands or rigid syntax to use Ask LLM. The MCP tools work via natural language — your AI client decides when to delegate and which provider to use.

Just Ask Naturally ​

Because Claude (and any other MCP-enabled assistant) natively integrates MCP tools, it knows when to route requests to Gemini, Codex, or Ollama based on what you say:

  • "Hey, can Gemini check if my config.js is valid?"
  • "Use Codex to suggest a better algorithm for @src/sort.ts"
  • "Ask Ollama to explain how this auth flow works (keep it local)."
  • "Have multi-llm send this question to Gemini and Codex in parallel — I want to compare their answers."
  • "What do Gemini and Codex think about this architectural decision? Give me their raw responses, no synthesis." → triggers /compare if you have the plugin
  • "Review my latest changes for security issues." → triggers /multi-review (verified findings) if you have the plugin

Mixing Tool Context Automatically ​

You can combine context (an error log, a stack trace, a diff) with a request without manually attaching files:

"I'm getting a null pointer error in my auth handler here. Have Gemini help me find the bug."

Your AI client extracts the relevant files from its conversation context and passes them to Gemini for you.

The @ File Syntax (Gemini) ​

When you want to explicitly include files in a prompt sent to Gemini, use the @ symbol:

text
Ask Gemini to summarize @README.md
Ask Gemini to review @src/auth.ts and @src/session.ts together
Ask Gemini to give me a high-level overview of @. (current directory)
Ask Gemini to scan @routes/**/*.js for OWASP issues

This is a Gemini CLI feature — @ syntax is interpreted by gemini, not by the MCP server. Codex and Ollama don't have direct equivalents (the relevant code should be quoted or pasted into the prompt).

Under the Hood — MCP Tools ​

For advanced users or when building automated AI workflows, these are the MCP tools the servers expose:

Unified orchestrator (ask-llm-mcp) ​

ask-llm ​

Send a prompt to any installed provider, picked via the provider parameter.

Parameters:

  • prompt (required): The question, code review request, or analysis task.
  • provider (required): One of gemini, codex, ollama (only providers detected at startup are accepted).
  • model (optional): Override the default model. Usually unnecessary — defaults are sensible per provider with auto-fallback.
  • sessionId (optional): Resume a previous conversation. Pass the value from a prior response's [Session ID: ...] or [Thread ID: ...] footer (or result.structuredContent.sessionId for programmatic clients).

Returns: Both human-readable text (content[0].text) AND a structured AskResponse (structuredContent) with {provider, response, model, sessionId, usage} — programmatic clients can extract fields directly without regex-parsing the footer.

multi-llm ​

Dispatch the same prompt to multiple providers in parallel; returns all responses in one structured payload.

Parameters:

  • prompt (required): The prompt to send to all selected providers.
  • providers (optional): Array of providers to dispatch to. Defaults to all available.

Returns: MultiLlmReport with {dispatchedAt, totalDurationMs, successCount, failureCount, results: [{provider, ok, response?, model?, sessionId?, usage?, durationMs, error?}, ...]}. Per-provider failures are isolated — one provider's quota issue doesn't fail the whole call. See ADR-066.

get-usage-stats ​

Per-session token totals, fallback counts, breakdowns by provider/model. In-memory, no persistence — resets when the MCP server restarts.

diagnose ​

Self-diagnosis: Node version, PATH resolution, provider CLI presence + versions. Read-only. Returns both human-readable text and a structured DiagnosticReport.

ping ​

Zero-cost connection test. Lists detected providers.

Per-provider servers (ask-gemini-mcp, ask-codex-mcp, ask-ollama-mcp) ​

Each per-provider server exposes its provider's ask-* tool with the richer per-provider parameter set, plus the shared get-usage-stats and ping.

ask-gemini ​

Same shape as ask-llm but always Gemini. Adds Gemini-specific behavior: @ file syntax, --include-directories support via includeDirs, stream-json live progressive output (ADR-057).

ask-gemini-edit ​

Returns structured OLD/NEW edit blocks rather than free-form text. Use this when you want Gemini to suggest specific code changes you can apply directly.

Parameters:

  • prompt (required): Describe the change you want.
  • model, includeDirs — same as ask-gemini.

fetch-chunk ​

Used automatically when Gemini's response is larger than a single MCP message allows. Returns subsequent chunks from the cached response.

ask-codex / ask-ollama ​

Same shape as ask-llm but pre-bound to the provider. Each accepts prompt, model, sessionId. Codex maps sessionId to its thread_id; Ollama uses server-side message replay.

MCP Resources ​

The orchestrator exposes one MCP Resource for live introspection:

  • usage://current-session — JSON snapshot of the in-memory SessionUsage accumulator. Read at any time for current totals. Same data as the get-usage-stats tool but accessible via resources/read instead of a tool call.

Plugin Slash Commands (Claude Code only) ​

If you've installed the Ask LLM plugin, additional slash commands are available:

  • /multi-review — parallel Gemini + Codex review with source verification of each finding
  • /gemini-review, /codex-review, /ollama-review — single-provider reviews
  • /brainstorm — multi-LLM brainstorm with Claude Opus as a first-class research participant
  • /compare — side-by-side responses, no synthesis (raw outputs)

See the Skills page for full descriptions.

Released under the MIT License.

Making AI collaboration simple, one tool at a time.