# Agent Mystro > An OpenAI-compatible router that sends each request to the model that has earned that kind of work, learned per customer from blind A/B picks. ## What it is A single endpoint (`/v1/chat/completions`) that classifies the incoming request into a task type, then selects among the customer's own provider accounts. Customers bring their own API keys and pay their providers directly; Agent Mystro charges for routing. ## What is measured - Cold drafting (battery b1, 10 briefs, blind, decoded 2026-07-25): google:pro 5, chain:claude_then_gpt 3, anthropic:opus 2, openai:flagship 0 - Web pages and interfaces (battery b2, 10 briefs, blind, decoded 2026-07-25): openai:flagship 8, google:pro 1, anthropic:opus 1 - Replies and threaded responses (battery b7, 10 briefs, blind, decoded 2026-07-25): chain:claude_then_gpt 6, google:pro 4, openai:flagship 0, anthropic:opus 0 - Live research (battery b8, 10 briefs, blind, decoded 2026-07-25): anthropic:opus 5, openai:flagship 3, google:pro 2 ## Multi-model creativity Models from different families working in concert measurably beat any single model on some work. The chain arm (Claude drafts, GPT edits), which won the reply battery 6 of 10 blind briefs against zero for either solo arm; POST /v1/council (mode rank), which fans 2-5 options to one seat per provider family on the customer's own keys and returns the ranking, the per-seat vote table, the strongest dissent, and an itemized per-seat cost; the same council in the browser at /app/council; and opt-in, revocable public sharing of a council result at /c/{token} (question, votes and verdict). ## How the learning works Bradley-Terry latent strengths per task type, one online gradient step per blind pick (learning rate 0.35), blended with the measured population prior by Bayesian shrinkage (kappa 6). ## Endpoints - POST /v1/chat/completions - OpenAI-compatible, streaming supported - POST /v1/council - rank 2-5 options across model families; vote table, dissent, per-seat cost - POST /v1/council/share - publish a council result at a public link (opt-in; cost stripped) - DELETE /v1/council/share/{token} - take a shared result offline - GET /v1/models - models reachable with the keys on file - POST /v1/keys - store a provider key (verified before storage) - POST /v1/account/rotate - rotate this account's own API key; the old key stops working immediately - POST /v1/taste/check - serve a blinded A/B pair - POST /v1/taste/pick - record a pick - GET /v1/taste/scorecard - your learned routing - GET /v1/routing/health - live model resolution and drift - GET /capabilities - full objective capability statement - GET /openapi.json - machine-readable API description - POST /mcp - MCP (Model Context Protocol) Streamable HTTP endpoint, stateless, JSON-RPC 2.0 ## MCP POST /mcp speaks MCP Streamable HTTP in stateless mode (no SSE stream, no session to resume; GET returns 405). Any MCP client (Claude Code, Codex CLI, ChatGPT desktop) can call it directly. Three tools, all gated behind the same bearer key as /v1: mystro_classify (task classification), mystro_route (advisory routing verdict, blends population priors with this account's own picks, calls no provider), and mystro_models (the current routing table). initialize and tools/list answer without a key (discovery); every tools/call needs Authorization: Bearer mk_live_....