# Model Maverick

> Model Maverick is a model recommendation engine for LLM workloads. It continuously benchmarks frontier models (Anthropic, OpenAI, Google, xAI at launch) against a private task corpus — 200+ tasks across ten capabilities and five difficulty tiers — and serves ranked, policy-aware model recommendations over REST and MCP. It performs no inference: applications take the recommendation and execute through their existing gateway or SDK.

Key facts:
- Answers one question: "which model should run this workload?"
- Nightly benchmarks on a private, versioned corpus; golden answers never enter a prompt (blocks contamination)
- Scores fold into a capability matrix: quality, pass rate, cost, and latency per (provider, model, endpoint, settings) column, smoothed with an exponentially weighted moving average (EWMA) so recent nights weigh most and one noisy night cannot flip a ranking
- Tenant policy is a hard boundary: named model pools, per-capability weights, cost/latency constraints; a request can never widen its own pool
- Every well-formed request returns a model — honest flags (measured, clears_difficulty, constraints_met) instead of refusals
- Full decision trace per recommendation; prompts are never stored (per-tenant keyed HMAC-SHA256 fingerprints only)
- The ranked list doubles as a fallback/escalation chain; each capability ships a machine-readable check spec
- Integration is one line: SDKs for TypeScript, Ruby, Go, and Python put the alias maverick/auto where a model name used to go (flagship hooks: Vercel AI SDK meta-provider, OpenAI/Anthropic SDK wrappers, RubyLLM, LiteLLM router strategy, Go http.RoundTripper); gateway plugins serve platform teams; the REST API stays open underneath
- Zero ceremony: outcome telemetry flows back automatically from SDKs and plugins (opt-out, and off means off) — customers never report anything
- Fail-open: an SDK failure can never block the customer's LLM call (cached chain → configured default → the model the code named)
- Execution backends are pluggable: keep per-provider keys and execute direct, or run the whole chain through one OpenAI-compatible backend key — OpenRouter first; Bedrock, Azure, and Vertex aliases use the same mechanism
- MCP server exposes recommend_model, get_policy, and list_models for agents (HTTP + stdio)
- Complements gateways (OpenRouter, LiteLLM, Bedrock, Vertex) and eval tools rather than replacing them
- Status: early access (July 2026)

## Pages

- [Home](https://modelmaverick.com/): What Model Maverick is and why hardcoded model names fail
- [How it works](https://modelmaverick.com/how-it-works/): The pipeline — capability matrix, grading, EWMA scoring, locked predictions, policy engine, decision traces
- [Capabilities](https://modelmaverick.com/capabilities/): The ten measured capabilities, difficulty tiers, and the high-stakes overlay
- [vs Routers & gateways](https://modelmaverick.com/vs/llm-routers-and-gateways/): Comparison with LLM gateways, routing APIs, and eval platforms
- [Early access](https://modelmaverick.com/early-access/): Design-partner program
- [Blog](https://modelmaverick.com/blog/): Announcements and engineering notes

## Contact

- Email: hello@modelmaverick.com