LLM Model Navigator MCP

An Apify standby MCP server that helps AI builders search the public OpenRouter model catalog, compare model metadata, estimate token costs, and make transparent heuristic choices. It uses the public endpoint through axios; no OpenRouter API key or other secret is required.

The streamable HTTP MCP endpoint is /mcp. Standby mode remains enabled, and GET / responds to Apify readiness probes.

Tools

search-llm-models

Search current catalog records by optional query (model ID, name, or description) and optional provider (the prefix before / in a model ID).

• query — optional non-empty string
• provider — optional non-empty string
• maxResults — optional integer, default 20, maximum 50

Returns each match's ID, name, provider, context length, modality, and prompt/completion prices per token when OpenRouter lists them.

compare-llm-models

Compare exact OpenRouter IDs in the order requested.

• modelIds — required array of 2 to 6 non-empty model IDs

Returns a structured comparison table with context length, input/output modalities, supported parameters, and listed prompt/completion per-token pricing. Missing IDs are explicitly reported rather than guessed.

estimate-llm-cost

Estimate a single request's text-token cost using OpenRouter's current listed pricing.

• modelId — required exact OpenRouter model ID
• inputTokens — required non-negative integer
• outputTokens — required non-negative integer

The result provides separate input and output USD estimates plus their total. It assumes the listed pricing.prompt and pricing.completion per-token values and excludes request, image, web-search, cache, internal-reasoning, tax, and other provider-specific charges. Models without both prices return an actionable error.

recommend-llm-model

Return up to five candidates for a workload.

• task — required: coding, reasoning, chat, vision, or cheap-batch
• maxInputTokens — optional minimum context length
• maxCostPerMillionTokens — optional cap on combined listed prompt + completion USD cost per 1M tokens

The recommendation score is deliberately transparent: task name/modality match, advertised context length, and lower listed combined prompt/completion prices. These are heuristics based on catalog metadata, not benchmark or quality claims. Validate quality, availability, and terms with your own workload before adopting a model.

Limits and reliability

• Results are derived from the live public OpenRouter catalog, so availability, capabilities, and prices can change.
• Search returns at most 50 models; comparison accepts 2–6 IDs; recommendations return at most 5 models.
• The service does not persist catalog data and does not use or accept OpenRouter secrets.
• If the catalog endpoint is unavailable or malformed, each tool returns a safe, actionable MCP error asking the caller to retry. It does not fabricate results.
• Each MCP tool call awaits Actor.charge({ eventName: 'tool-call' }) for Apify pay-per-event billing.

Local development

npm install
npm run build
npm test
apify validate-schema
npm run start:dev

The server listens on port 3000 by default, or APIFY_CONTAINER_PORT when set. Connect an MCP client to http://localhost:3000/mcp.

Deployment

Deploy with Apify in standby mode and use the actor's /mcp endpoint. The actor endpoint itself is authenticated by Apify as appropriate for your deployment; this service makes unauthenticated public requests only to OpenRouter's model-catalog endpoint.