OpenRouter

Unified, OpenAI-compatible model routing API. One endpoint reaches many model providers, with provider fallbacks, cost/performance routing, app attribution, and router models such as openrouter/auto, openrouter/free, and OpenRouter Fusion. Source: OpenRouter quickstart, 2026-06-17

OpenRouter is the model-layer abstraction Kevin keeps using as an analogy for agent infrastructure: one API key and one request shape over many model providers. It supports direct REST calls, OpenAI SDK compatibility by changing the base URL, generated client SDKs, and an Agent SDK. Source: OpenRouter quickstart, 2026-06-17

Routing Behavior

For a specific model slug, OpenRouter selects an available provider endpoint and falls back when a provider returns server errors or rate limits. Provider routing can prioritize price by default or use caller constraints for throughput, required capabilities, quantization, region, or provider allow/deny lists. Tool use filters to providers that support tool calling. Source: OpenRouter API reference, 2026-06-17; Source: Provider routing docs, 2026-06-17

OpenRouter also exposes router models:

MCP Development Assistant

OpenRouter's remote MCP server connects coding agents to live model data, pricing, benchmarks, docs search, credits, provider lists, app rankings, bundled skill files, and quick test inference through https://mcp.openrouter.ai/mcp. The setup is one remote HTTP MCP URL plus OAuth login; OpenRouter says the login mints a dedicated API key with a 7-day expiry and a default $10 spend cap that can be edited or revoked. A direct unauthenticated request checked on 2026-07-03 returned 401 invalid_token with a WWW-Authenticate OAuth-protected-resource header, matching the docs' auth model. The docs were rechecked on 2026-07-04 and still frame the MCP as a build-time assistant; production apps should call the OpenRouter API directly. Source: OpenRouter MCP docs, 2026-07-04; Source: OpenRouter MCP announcement, 2026-06-25

Treat the MCP as a development assistant, not the production inference path. Most tools are read-only lookups; chat-send is the exception because it sends an explicit prompt to a model and spends against the MCP key balance. Apps should still call the OpenRouter API directly for production inference. Source: OpenRouter MCP announcement FAQ, 2026-06-25

OpenRouter's own model-council example adds the operating constraint: use live MCP data to tune the council to the current codebase and use case. That means the MCP is a selection/evidence tool for model roles, pricing, capabilities, and test prompts, while the durable council recipe lives in LLM Council. Source: X/@OpenRouter, 2026-06-26

Open-Weight Agentic Models

OpenRouter's June 2026 open-weight analysis says open-weight models have stayed roughly within a 3-6 month capability gap of frontier labs while becoming cost-attractive for scaled agentic workloads. The fetched source names DeepSeek V4 Flash as the first open-weight model to cross an "agentic rubicon," GLM-5.2 as the open model that makes Opus-style agentic coding portable, MiniMax M3 as the long-context multimodal lane for screenshot, UI, diagram, video, and document-heavy agents, and NVIDIA Nemotron 3 Ultra as the U.S.-built open-weight option when vendor/deployment story matters. Source: OpenRouter Insights, 2026-06-27

Use the comparison card as a routing prior, not a fixed default:

Pricing, throughput, and provider quality should be checked in OpenRouter at selection time; the durable decision is the task-to-model lane, not the June 2026 number itself. Source: OpenRouter Insights, 2026-06-27; Source: local comparison card review, 2026-07-03

Stack Fit

• Agent Machines analogy: Agent Machines is "OpenRouter for agents and containers": runtime × substrate × model routing, not only model routing. Source: Agent Machines
• App integration: Vercel AI SDK / AI Gateway stay the default for Vercel apps; OpenRouter is useful when its model catalog, provider routing, or Fusion surface wins.
• Model evaluation: OpenRouter lowers switching cost, but it does not remove the need for The Eval Loop (Slop Is an Output Problem) checks. A model/router swap without an eval is just vibes with billing.

---

Timeline

• 2026-07-04 | Reviewed OpenRouter's first-party MCP model-council example and rechecked official MCP docs. Kept MCP as a build-time model-selection assistant with live catalog/pricing/benchmark/docs/test-call tools, while production inference remains direct API/gateway work. Source: X/@OpenRouter, 2026-06-26; Source: OpenRouter MCP docs
• 2026-07-03 | Deep-reviewed the OpenRouter comparison card and expanded the June 2026 open-weight model note to include Nemotron 3 Ultra plus a task-to-model routing table. Source: X/@OpenRouter, 2026-06-27; Source: OpenRouter Insights
• 2026-06-17 | Page created during the full wiki/tool-routing audit because OpenRouter Fusion, Agent Machines, Loop, and model-routing docs all referenced OpenRouter without a canonical tool page. Source: User, 2026-06-17
• 2026-06-28 | Added OpenRouter's June 2026 open-weight model analysis as model-routing evidence for agentic workloads. Source: X/@OpenRouter, 2026-06-27; Source: OpenRouter Insights
• 2026-07-03 | Added the OpenRouter MCP server as a coding-agent development assistant for live model catalog, pricing, benchmarks, docs search, credits, provider lists, and test inference; kept production apps routed through the OpenRouter API. Source: X/@lennox_saint, 2026-06-26; Source: OpenRouter MCP docs