> ## Documentation Index > Fetch the complete documentation index at: https://docs.reasonblocks.com/llms.txt > Use this file to discover all available pages before exploring further. # Endpoint compatibility > LLM API paths, client SDKs, upstream routing, and which features apply per endpoint on the ReasonBlocks gateway. The hosted gateway is a **byte-transparent reverse proxy**. Your client sends normal provider HTTP requests; ReasonBlocks forwards them to the configured upstream with the same path and query string. There is no fixed allowlist of routes. **Default product path:** Anthropic Messages (`POST /v1/messages`) via `ANTHROPIC_BASE_URL`. That path has the best trace reconstruction, supports training, and **relegation serving** (cheaper model first, frontier fallback). It matches how most agent frameworks call Claude today. Other common endpoints (OpenAI Chat Completions, Responses API, Gemini, etc.) work for **capture and passthrough** when upstream routing is set correctly. Relegation serving (cheaper model + frontier fallback) applies only to `POST /v1/messages`. ## Upstream routing (operator / hosted config) ReasonBlocks sets upstream routing on hosted gateways. Self-hosters configure: | Variable | Default | Role | | -------------------------- | --------------------------- | --------------------------------------- | | `AGENTRECON_UPSTREAM_BASE` | `https://api.anthropic.com` | Base URL; client path appended verbatim | | `AGENTRECON_PROVIDER` | `anthropic` | Label on capture metadata only | To forward OpenAI-shaped clients, point upstream at OpenAI: ```bash theme={null} export AGENTRECON_UPSTREAM_BASE=https://api.openai.com ``` Hosted tenants: use the dashboard gateway URL; do not change upstream unless your account supports multi-provider routing there. ## Common LLM API paths Paths the optional client shim treats as LLM calls (for `X-RB-Run` / `X-RB-Seq` headers). All other paths still forward through the catch-all proxy. | Client path | Typical upstream | Agent env var | | --------------------------------------- | ------------------------------ | --------------------------- | | `POST /v1/messages` | Anthropic API | `ANTHROPIC_BASE_URL` | | `POST /v1/chat/completions` | OpenAI API | `OPENAI_BASE_URL` | | `POST /v1/responses` | OpenAI API | `OPENAI_BASE_URL` | | `POST /v1/complete` | OpenAI API | `OPENAI_BASE_URL` | | `POST …:generateContent` | Google Generative Language API | Provider base URL + gateway | | Other (`/v1/models`, Azure paths, etc.) | Your provider | Base URL pointed at gateway | ### Anthropic (default) ```bash theme={null} export ANTHROPIC_BASE_URL="https://gateway..reasonblocks.com" export ANTHROPIC_API_KEY="sk-ant-..." # passthrough; never stored python your_agent.py ``` Works with the Anthropic SDK, LangChain, CrewAI, and any client that respects `ANTHROPIC_BASE_URL`. ### OpenAI Chat Completions Gateway upstream must be OpenAI (hosted config or self-hosted `AGENTRECON_UPSTREAM_BASE`): ```bash theme={null} export OPENAI_BASE_URL="https://gateway..reasonblocks.com" export OPENAI_API_KEY="sk-..." # passthrough; never stored ``` Client calls `POST /v1/chat/completions` on the gateway; traffic forwards to OpenAI. Capture works; structured golden traces and relegation serving are **weaker or unavailable** on this path (see below). ### OpenAI Responses API Same pattern as Chat Completions: `OPENAI_BASE_URL` at the gateway, upstream `https://api.openai.com`, path `/v1/responses`. ### Gemini and other providers Point the client base URL at the gateway and set `AGENTRECON_UPSTREAM_BASE` to the provider origin (for example `https://generativelanguage.googleapis.com`). URLs containing `:generateContent` are recognized by the client shim for run headers. ## Framework compatibility | Client | Configuration | | ---------------------------- | --------------------------------------------- | | Anthropic SDK (Python/TS) | `ANTHROPIC_BASE_URL` → gateway | | LangChain | Inherits `ANTHROPIC_BASE_URL` | | CrewAI | Inherits `ANTHROPIC_BASE_URL` | | OpenAI SDK | `OPENAI_BASE_URL` → gateway + OpenAI upstream | | LiteLLM / httpx-based stacks | Base URL override to gateway | | Custom HTTP | Any method/path the catch-all accepts | API keys pass through upstream and are **not stored** by ReasonBlocks. ## Feature matrix by endpoint | Feature | `POST /v1/messages` | OpenAI-shaped paths | Gemini / other | | ---------------------------- | --------------------------------------- | ------------------------------ | --------------- | | Raw capture | Yes | Yes | Yes | | Structured reconstruction | Best | Partial | Sparse | | Golden runs for distillation | Yes (Tier B/C recommended) | Limited | Not recommended | | Relegation serving | Yes (cheaper model + frontier fallback) | No (upstream passthrough only) | No | Training and relegation serving use flat Anthropic-shaped transcripts. If your workflow must use the cheaper model path, run the agent on **`/v1/messages`** even if you also capture other provider traffic. ## Relegation serving (Anthropic path only) After training, keep the same gateway URL. Relegation routing affects **only** `POST /v1/messages`: | Routing | Behavior | | --------------------- | ------------------------------------------------------------- | | **Cheaper model** | Trained task-specific model handles the step first | | **Frontier fallback** | Frontier model when the cheaper model lacks coverage or fails | | **Edge cases** | Fallback traffic is captured for the next training cycle | OpenAI-shaped requests are not rewritten to the cheaper model. They passthrough to `AGENTRECON_UPSTREAM_BASE`. Enable relegation in the dashboard (hosted) or via gateway config (self-hosted). See [Quickstart](/quickstart) steps 5–6. ## Client shim paths When using [Tier B or C](/client-integration), `rbtrace.client.install()` stamps run headers on requests matching: * `/v1/messages` * `/v1/chat/completions` * `/v1/responses` * `/v1/complete` * paths containing `:generateContent` Install: `pip install rbtrace` (when published to PyPI). ## Gateway admin endpoints Not forwarded to upstream: | Method | Path | Purpose | | ------------- | ------------------------ | ------------------------------------ | | `GET` | `/__rbtrace/stats` | Health, queue depth, exchange counts | | `POST`, `GET` | `/__rbtrace/flush` | Flush batched writes before backup | | `POST` | `/__rbtrace/writer/kill` | Fail-open testing | Self-hosted operators: see [Deployment](/deployment#advanced-self-hosting-the-gateway). ## Limitations * Upstream failures return **502** with Anthropic-shaped error JSON regardless of client type. * Cheaper-model responses are synthesized Anthropic Messages JSON (or a single SSE burst), not native provider streaming. * Integration tests in the gateway codebase focus on `POST /v1/messages`; other paths are supported by architecture but less validated end-to-end. ## Related Tier A/B/C and the optional shim. Env vars and serve contract.