Consult LLM MCP

Instructions

Implementation Reference

• src/server.ts:68-157 (handler)

  Main handler function for the 'consult_llm' tool. Parses arguments using ConsultLlmArgs schema, determines model, builds prompt with files and git diff context, handles web_mode by copying to clipboard, otherwise queries the LLM via queryLlm and returns the response.
  export async function handleConsultLlm(args: unknown) { const parseResult = ConsultLlmArgs.safeParse(args) if (!parseResult.success) { const errors = parseResult.error.issues .map((issue) => `${issue.path.join('.')}: ${issue.message}`) .join(', ') throw new Error(`Invalid request parameters: ${errors}`) } const { files, prompt: userPrompt, git_diff, web_mode, model: parsedModel, } = parseResult.data const providedModel = typeof args === 'object' && args !== null && Object.prototype.hasOwnProperty.call( args as Record<string, unknown>, 'model', ) const model: SupportedChatModel = providedModel ? parsedModel : (config.defaultModel ?? parsedModel) logToolCall('consult_llm', args) const isCliMode = isCliExecution(model) let prompt: string let filePaths: string[] | undefined if (web_mode || !isCliMode) { const contextFiles = files ? processFiles(files) : [] const gitDiffOutput = git_diff ? generateGitDiff(git_diff.repo_path, git_diff.files, git_diff.base_ref) : undefined prompt = buildPrompt(userPrompt, contextFiles, gitDiffOutput) } else { filePaths = files ? files.map((f) => resolve(f)) : undefined const gitDiffOutput = git_diff ? generateGitDiff(git_diff.repo_path, git_diff.files, git_diff.base_ref) : undefined prompt = gitDiffOutput ? `## Git Diff\n\`\`\`diff\n${gitDiffOutput}\n\`\`\`\n\n${userPrompt}` : userPrompt } await logPrompt(model, prompt) if (web_mode) { const systemPrompt = getSystemPrompt(isCliMode) const fullPrompt = `# System Prompt ${systemPrompt} # User Prompt ${prompt}` await copyToClipboard(fullPrompt) let responseMessage = '✓ Prompt copied to clipboard!\n\n' responseMessage += 'Please paste it into your browser-based LLM service and share the response here before I proceed with any implementation.' if (filePaths && filePaths.length > 0) { responseMessage += `\n\nNote: File paths were included:\n${filePaths.map((p) => ` - ${p}`).join('\n')}` } return { content: [{ type: 'text', text: responseMessage }], } } const { response, costInfo } = await queryLlm(prompt, model, filePaths) await logResponse(model, response, costInfo) return { content: [{ type: 'text', text: response }], } }
• src/server.ts:159-171 (registration)

  Registers the CallToolRequest handler, which dispatches to handleConsultLlm when the tool name is 'consult_llm'.
  server.setRequestHandler(CallToolRequestSchema, async (request) => { if (request.params.name === 'consult_llm') { try { return await handleConsultLlm(request.params.arguments) } catch (error) { throw new Error( `LLM query failed: ${error instanceof Error ? error.message : String(error)}`, ) } } throw new Error(`Unknown tool: ${request.params.name}`) })
• src/server.ts:49-53 (registration)

  Registers the ListToolsRequest handler, which returns the schema for the 'consult_llm' tool.
  server.setRequestHandler(ListToolsRequestSchema, () => { return { tools: [toolSchema], } })
• src/schema.ts:16-64 (schema)

  Zod schema defining the input arguments for the 'consult_llm' tool, including files, prompt, model, web_mode, and git_diff options.
  export const ConsultLlmArgs = z.object({ files: z .array(z.string()) .optional() .describe( 'Array of file paths to include as context. All files are added as context with file paths and code blocks.', ), prompt: z .string() .describe( 'Your question or request for the consultant LLM. Ask neutral, open-ended questions without suggesting specific solutions to avoid biasing the analysis.', ), model: SupportedChatModel.optional() .default('o3') .describe( 'LLM model to use. Prefer gpt-5.1-codex-max when user mentions Codex. This parameter is ignored when `web_mode` is `true`.', ), web_mode: z .boolean() .optional() .default(false) .describe( "If true, copy the formatted prompt to the clipboard instead of querying an LLM. When true, the `model` parameter is ignored. Use this to paste the prompt into browser-based LLM services. IMPORTANT: Only use this when the user specifically requests it. When true, wait for the user to provide the external LLM's response before proceeding with any implementation.", ), git_diff: z .object({ repo_path: z .string() .optional() .describe( 'Path to git repository (defaults to current working directory)', ), files: z .array(z.string()) .min(1, 'At least one file is required for git diff') .describe('Specific files to include in diff'), base_ref: z .string() .optional() .default('HEAD') .describe( 'Git reference to compare against (e.g., "HEAD", "main", commit hash)', ), }) .optional() .describe( 'Generate git diff output to include as context. Shows uncommitted changes by default.', ), })
• src/schema.ts:66-78 (schema)

  Defines the JSON schema and metadata (name, description, inputSchema) for the 'consult_llm' tool, used in tool listing.
  const consultLlmInputSchema = z.toJSONSchema(ConsultLlmArgs, { target: 'openapi-3.0', }) export const toolSchema = { name: 'consult_llm', description: `Ask a more powerful AI for help with complex problems. Provide your question in the prompt field and always include relevant code files as context. Be specific about what you want: code implementation, code review, bug analysis, architecture advice, etc. IMPORTANT: Ask neutral, open-ended questions. Avoid suggesting specific solutions or alternatives in your prompt as this can bias the analysis. Instead of "Should I use X or Y approach?", ask "What's the best approach for this problem?" Let the consultant LLM provide unbiased recommendations.`, inputSchema: consultLlmInputSchema, } as const