Anthropic API

The endpoint

POST https://api.experientiallabs.ai/v1/messages speaks the Anthropic Messages wire protocol. Authenticate with the Anthropic-style x-api-key header or with Authorization: Bearer, either carries the same xpl_ key. Name any slug from GET https://api.experientiallabs.ai/v1/models as the model.

curl "https://api.experientiallabs.ai/v1/messages" \
  -H "x-api-key: $EXPLABS_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{"model": "qwen3.8-27b", "max_tokens": 1024, "messages": [{"role": "user", "content": "Hello"}]}'

Streaming

Set stream: trueto receive Anthropic's server-sent event stream (message_start, content_block_delta, message_stop, and the rest), exactly as the Anthropic SDKs expect.

curl -N "https://api.experientiallabs.ai/v1/messages" \
  -H "x-api-key: $EXPLABS_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{"model": "qwen3.8-27b", "max_tokens": 1024, "stream": true, "messages": [{"role": "user", "content": "Hello"}]}'

Retries and route selection

Messages generation accepts the same gateway.retry and gateway.routing extension as Chat and Responses. Python SDK users pass it through extra_body; disable the SDK's own retries independently withmax_retries=0. This example allows exactly one gateway generation dispatch, with no fallback or hidden repair redial.

curl "https://api.experientiallabs.ai/v1/messages" \
  -H "x-api-key: $EXPLABS_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
  "model": "qwen3.8-27b",
  "max_tokens": 1024,
  "messages": [
    {
      "role": "user",
      "content": "Hello"
    }
  ],
  "gateway": {
    "retry": {
      "max_attempts_per_route": 1,
      "max_total_attempts": 1,
      "backoff": {
        "type": "none"
      }
    },
    "routing": {
      "allow_fallbacks": false
    }
  }
}'

For bounded retries, configurable exponential backoff, or selecting an eligible lower rung by its opaque route_id, see the request policy guide. The extension is not valid on /v1/messages/count_tokens or batch lines. A Messages Idempotency-Key still does not create replay protection.

Translation lane limits

The lane maps onto the gateway's shared chat surface, so a few Anthropic features are route-scoped. Where a route cannot express something, the gateway translates or drops it with a disclosure instead of failing the request; only signed provider state is rejected:

• Extended thinking works on all-Anthropic routes: the thinkingconfig passes through verbatim and thinking/redacted-thinking history blocks round-trip with signatures intact. On a non-Anthropic reasoning route the config translates to the route's nearest reasoning effort (disclosed as thinking->reasoning_effort:<tier>; a bare thinking: {type: enabled} is accepted and reads as the default depth; an explicit output_config.effort wins with thinking->dropped(superseded_by_effort)), and on a route with no reasoning it is dropped with a disclosure. Block-level cache_control markers are preserved on supported Anthropic and OpenRouter adapters and translated into Bedrock cache checkpoints. Generic adapters that cannot carry a marker disclose it as not forwarded; a successful request or a listed cached price does not guarantee a cache hit. There is no organization caching switch. Platform-funded requests support five-minute cache writes; one-hour markers are refused until duration-specific pricing is available. Usage distinguishes cache_read_input_tokens and cache_creation_input_tokens. OpenRouter's reasoning object (effort or max_tokens, plus enabled) is accepted beside thinkingand wins when both are present. On a reasoning-exposed non-Anthropic rung the model's own reasoning streams as an unsigned thinking block (tool turns carry it sealed in a trailing redacted_thinking block) and replays on the next turn; Anthropic-signed thinking blocks replayed onto a non-Anthropic route are dropped with the messages.thinking->dropped(unsupported_by_provider) disclosure instead of a 400.
• Image blocks are served on every route whose model accepts images, up to 100 per request. Document (PDF) blocks need a document-capable route; otherwise a 400 names messages and suggests a capable alias.
• tool_result.is_error=true is native on Anthropic rungs and folds into the result text on every other wire (disclosed), so a failed tool call never ends a session.
• Idempotency-Key is not honored on this lane (Anthropic defines none).
• /v1/messages/count_tokensanswers with the gateway's own token estimate for the route (the response discloses that it is an estimate, not the provider's count); per-turn usage on the stream is the authoritative figure.

Claude Code and other agents

Claude Code connects through this same lane by pointing ANTHROPIC_BASE_URL at the gateway and passing an xpl_ key as ANTHROPIC_API_KEY. Its startup connectivity probe, a keyless HEAD /api/hello, answers 200 exactly as Anthropic's API does; it is not a request and needs no key. The per-agent configuration is in Coding agents.

See also

Prefer the OpenAI protocol? The Quickstart covers Chat Completions and the Responses API. The full surface is in the API reference.