MCP Ollama Consult Server

Instructions

Implementation Reference

• src/handlers.ts:221-269 (handler)

  Main handler logic for 'compare_ollama_models': validates input, resolves models with fallbacks if provided or uses top 2 available, generates responses from each via Ollama API, collects side-by-side results or errors.
  case "compare_ollama_models": { let modelsArg = args?.models as string[] | undefined; const prompt = args?.prompt as string; const system_prompt = args?.system_prompt as string | undefined; if (!prompt) { return { content: [ { type: "text", text: "Missing required argument: prompt", }, ], isError: true, }; } let modelsToUse: string[] = []; if (Array.isArray(modelsArg) && modelsArg.length > 0) { // Resolve each requested model with fallback const resolved = await Promise.all(modelsArg.map((m) => resolveModel(m))); modelsToUse = resolved.map((r) => r.model); } else { try { const resp = await axios.get(`${OLLAMA_BASE_URL}/api/tags`); modelsToUse = (resp.data.models || []).map((m: any) => m.name).slice(0, 2); } catch (err) { modelsToUse = ["llama2"]; } } const contents: any[] = []; for (const m of modelsToUse) { try { const gen = await axios.post(`${OLLAMA_BASE_URL}/api/generate`, { model: m, prompt, system: system_prompt, stream: false, }); contents.push({ type: "text", text: `Model ${m}:\n${gen.data.response}` }); } catch (e) { const message = e instanceof Error ? e.message : String(e); contents.push({ type: "text", text: `Model ${m} failed: ${message}` }); } } return { content: contents }; }
• src/handlers.ts:112-125 (registration)

  Tool registration in listTools(): defines name, description, and input schema for compare_ollama_models.
  { name: "compare_ollama_models", description: "Run the same prompt against multiple Ollama models and return their outputs side-by-side for comparison. Requested models that are unavailable automatically fall back to cloud models or local alternatives. Handles unavailable models gracefully without breaking the comparison.", inputSchema: { type: "object", properties: { models: { type: "array", items: { type: "string" } }, prompt: { type: "string" }, system_prompt: { type: "string" }, }, required: ["prompt"], }, },
• src/handlers.ts:116-124 (schema)

  Input schema defining parameters: models (array of strings, optional), prompt (string, required), system_prompt (string, optional).
  inputSchema: { type: "object", properties: { models: { type: "array", items: { type: "string" } }, prompt: { type: "string" }, system_prompt: { type: "string" }, }, required: ["prompt"], },
• src/handlers.ts:50-88 (helper)

  Helper function resolveModel used by the handler to automatically fallback unavailable models to cloud (e.g., deepseek-v3.1:671b-cloud) or local alternatives like mistral.
  async function resolveModel(requestedModel: string): Promise<{ model: string; isCloud: boolean; fallback: boolean; }> { // If it's explicitly a cloud model, use it if (requestedModel.includes(":cloud")) { return { model: requestedModel, isCloud: true, fallback: false }; } // Check if it's available locally const available = await getAvailableLocalModels(); if (available.includes(requestedModel)) { return { model: requestedModel, isCloud: false, fallback: false }; } // Not available locally - try cloud models for (const cloudModel of CLOUD_MODELS) { // In real implementation, you might verify cloud models are actually available // For now, we assume they are return { model: cloudModel, isCloud: true, fallback: true }; } // If cloud fails, fall back to local models for (const localFallback of LOCAL_FALLBACKS) { if (available.includes(localFallback)) { return { model: localFallback, isCloud: false, fallback: true }; } } // Last resort: use whatever is available, or the original request if (available.length > 0) { return { model: available[0], isCloud: false, fallback: true }; } // If nothing is available, return the requested model anyway // (will fail gracefully in the actual API call) return { model: requestedModel, isCloud: false, fallback: true }; }
• src/handlers.ts:28-44 (helper)

  Cached helper to fetch list of available local Ollama models from /api/tags endpoint.
  async function getAvailableLocalModels(): Promise<string[]> { const now = Date.now(); if (availableModelsCache && now - cacheTimestamp < CACHE_TTL) { return availableModelsCache; } try { const response = await axios.get(`${OLLAMA_BASE_URL}/api/tags`); const models = (response.data.models || []).map((m: any) => m.name); availableModelsCache = models; cacheTimestamp = now; return models; } catch { // If we can't reach Ollama, return empty (we'll fall back to cloud) return []; } }