# Compatibility matrix What OpenSmartRoute connects to, how each connection is made, and how it was verified. One row per provider, client, framework and adapter; the *Verified* column names the evidence - a unit test that runs on every push, a live run with real credentials and its date, or "by protocol" when the connection is the OpenAI-compatible wire format and no deeper adapter exists. A test holds this page to the code: every provider preset the platform offers, every adapter module of the SDK and every framework the website names must have a row here. Legend for *Verified*: **CI** = exercised by the test suite on every push (a fake endpoint speaking the provider's wire format); **live ``** = a real call with real credentials on that date, recorded in the repository; **protocol** = reached through the OpenAI-compatible API without a dedicated adapter, so what is verified is the protocol, not the vendor's quirks. ## Model providers (hosted platform and `osr serve`) The platform's *Model providers* page offers these presets (`platform/api/osr_platform/providers.py`); each is an OpenAI-compatible endpoint the router calls with the workspace's own key. The SDK reaches the same endpoints through `attach_chat_handlers` / `OpenAICompatClient` and, for Anthropic, `AnthropicClient`. | Preset | Vendor API | Wire format | Verified | |---|---|---|---| | `openai` | OpenAI | OpenAI chat / embeddings / responses | CI; live 2026-09-10 (token calibration against billed gpt-4.1 usage) | | `azure` | Azure OpenAI / Azure AI Foundry | OpenAI v1 with `api-key` header | CI; live 2026-09-10 (multi-cloud benchmark, 150 prompts) | | `azure-identity` | Azure OpenAI / Azure AI Foundry, keyless | OpenAI v1 with a managed-identity bearer token (`identity://https://cognitiveservices.azure.com`) | CI (reference resolution and caching); the token exchange is the same identity endpoint the Key Vault references use in production | | `huggingface` | Hugging Face Inference Providers | OpenAI-compatible router (`router.huggingface.co/v1`, Hub model ids) | protocol | | `bedrock` | Amazon Bedrock | OpenAI-compatible endpoint (`openai.gpt-oss-*`) | live 2026-09-10 (multi-cloud benchmark) | | `gemini` | Google Gemini API | OpenAI-compatible endpoint | protocol | | `vertex` | Google Cloud Vertex AI | OpenAI-compatible endpoint (`google.thinking_config` passed through) | live 2026-09-10 (multi-cloud benchmark, judge re-scoring) | | `openrouter` | OpenRouter | OpenAI | CI (the example catalogue names OpenRouter ids) | | `anthropic` | Anthropic | native Messages API (`AnthropicClient`) | CI (translation of chat calls, streaming) | | `mistral` | Mistral | OpenAI | protocol | | `groq` | Groq | OpenAI | protocol | | `together` | Together AI | OpenAI | protocol | | `fireworks` | Fireworks AI | OpenAI | protocol | | `deepseek` | DeepSeek | OpenAI | protocol | | `xai` | xAI | OpenAI | protocol | | `cohere` | Cohere | OpenAI-compatible endpoint | protocol | | `ollama` | Ollama (local models) | OpenAI | CI; live, in production (the on-prem targets of opensmartroute.ai run on it) | | `vllm` | vLLM | OpenAI | protocol (the vLLM semantic-router import is CI) | | `litellm` | LiteLLM proxy | OpenAI | protocol | | `custom` | any OpenAI-compatible endpoint | OpenAI | CI (the test fleet is one) | The platform's own model server (`osr-models`, Ollama and OpenAI APIs over llama.cpp) is a `ollama` preset to the router; its API coverage has its own matrix in the model-serving plan. ## Clients that talk to the router | Client | How | Verified | |---|---|---| | OpenAI Python SDK | `OpenAI(base_url=/v1, api_key=)`, `model="auto"` | CI (the official client against `osr serve`, chat, streaming, tools, embeddings); `openai>=1.50,<3` | | OpenAI Node SDK and every OpenAI-compatible client (Vercel AI SDK, Continue, aider, LiteLLM, ...) | same base URL | protocol; the TypeScript SDK's live test runs against production | | Anthropic SDK / Claude Code | `/v1/messages` with the router as `base_url` | CI (Messages API surface) | | OpenAI Responses clients | `/v1/responses` | CI | | MCP hosts (VS Code, Cursor, Windsurf, Claude Desktop, Claude Code) | `/mcp` (Streamable HTTP) or `osr mcp` (stdio) | CI (the official `mcp` client, `mcp>=1.10,<2`); live in production | | `@opensmartroute/sdk` (TypeScript) | typed REST client over the platform API | CI (surface and live tests) | | `osr` CLI | the SDK, in a terminal | CI | ## Agent frameworks | Framework | Integration | Depth | Verified | |---|---|---|---| | LangGraph | `adapters.frameworks.langgraph_node` / `langgraph_condition` | deep adapter (a routing node and an edge condition) | CI (the node runs without LangGraph installed; the graph wiring is the framework's) | | Microsoft Agent Framework / AutoGen | `adapters.frameworks.maf_router_executor` | deep adapter (an executor) | CI | | OpenAI function calling / Agents SDK | `adapters.frameworks.openai_tool_spec` / `openai_tool_handler` | deep adapter (the router as a tool) | CI | | LangChain | `ChatOpenAI(base_url=/v1, model="auto")` | protocol | protocol | | CrewAI | the OpenAI-compatible LLM setting | protocol | protocol | | LlamaIndex | `OpenAILike(api_base=/v1)` | protocol | protocol | | Semantic Kernel | the OpenAI connector with a base URL | protocol | protocol | | pydantic-ai | the OpenAI model with a base URL | protocol | protocol | ## SDK adapters (`opensmartroute.adapters`) | Module | Connects | Verified | |---|---|---| | `a2a` | A2A Agent Cards as agent targets; `a2a_handler` with task polling | CI | | `anthropic_client` | the Anthropic Messages API as a chat handler | CI | | `catalogue` | the live model catalogue (prices, context windows) | CI; live every 8 hours on the platform | | `frameworks` | LangGraph, Agent Framework, OpenAI tools (above) | CI | | `handlers` | HTTP endpoints, MCP tools and asynchronous queues as handlers | CI | | `harness` | agent harnesses: callable, HTTP, subprocess | CI | | `marketplace` | the marketplace (search, pull, sync, signed manifests) | CI; live in production | | `mcp` | MCP tool catalogues (stdio and Streamable HTTP), resources and prompts | CI (official client) | | `mcp_servers` | MCP server recommendation from a request | CI | | `mock_client` | the `mock` provider kind: a deterministic, network-free provider for tests, demos and pipelines | CI | | `openai_compat` | OpenAI-compatible HTTP client (chat, embeddings, retries, error mapping) | CI; live (above) | | `optional` | adapters with extra dependencies, imported lazily | CI | | `personas` | persona catalogues as targets | CI | | `semantic_router` | a vLLM semantic-router configuration imported as rules | CI | | `skills` | Agent-Skills `SKILL.md` packages as targets | CI | | `websearch` | web knowledge and feeds for the self-improver | CI; live on the platform | ## Runtimes | Runtime | Verified | |---|---| | CPython 3.11, 3.12 (Linux, Windows) | CI on every push; 3.11 is the floor | | Node 22 (web, TypeScript SDK, extension, desktop) | CI | | PostgreSQL 16 (platform) | CI and production | | Redis, Kafka (platform clustering, optional) | CI (platform cluster job) | A vendor that changes its wire format breaks a *protocol* row first; the CI rows carry a fake of the format the SDK was written against, so a change shows up as a red build only once the fake is updated - treat a *protocol* row as "should work" and a *CI* or *live* row as "does work".