API Quickstart

cURL

Zero-dependency smoke test against the chat completions endpoint. Set $INFERWAY_API_KEY to a gw_live_… direct key (create one in the Console) in your shell, paste, and run.

curl https://api.inferway.ai/v1/chat/completions \
  -H "Authorization: Bearer $INFERWAY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gwmm/gemma-4-12b-it",
    "messages": [{"role": "user", "content": "Say hello in one short sentence."}],
    "max_tokens": 64
  }'

Python (openai SDK)

The official OpenAI Python SDK works against any OpenAI-compatible base URL. Swap the base URL to Inferway, keep your existing direct key.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.inferway.ai/v1",
    api_key="$INFERWAY_API_KEY",
)

resp = client.chat.completions.create(
    model="gwmm/gemma-4-12b-it",
    messages=[{"role": "user", "content": "Say hello in one short sentence."}],
    max_tokens=64,
)

print(resp.choices[0].message.content)

Endpoint reference

Error codes

Inferway returns OpenAI-compatible error objects. Example shape: &123;"error": &123; "type": "...", "message": "...", "code": "..." &125;&125;

Rate limits

Inferway applies transparent per-subject rate limits. Every /v1/* inference response includes standard X-RateLimit-* headers. When a limit is hit, the response is HTTP 429 with a Retry-After header and a rate_limited error object pointing to the rate-limit docs.

• X-RateLimit-Limit. The quota of the current RPM window (5 for guests, 60 for API keys).
• X-RateLimit-Remaining. Requests remaining in the current RPM window.
• X-RateLimit-Reset. Unix timestamp when the current RPM window resets.
• Retry-After. Seconds to wait before retrying after a 429 response.

Operational limits

• Concurrency ceiling. 64 simultaneous in-flight requests per API key. Excess returns HTTP 429.
• First-token timeout. 20 s.
• Context window. 28K tokens (input + output combined).
• Precision. FP8, served at advertised precision under load.
• Data retention. Zero. See Privacy Policy for the full architecture.
• Authentication. Direct access requires a gw_live_… / gw_test_… key (issued via the Console). OpenRouter customers should point at https://openrouter.ai/api/v1 with their OpenRouter key instead.

MCP (Model Context Protocol)

Inferway also exposes an MCP server so AI agents can call it natively. The MCP endpoint is https://api.inferway.ai/v1/mcp/sse (HTTP/SSE transport).

Available tools:

• chat_completion — run a non-streaming chat completion
• list_models — list models, context windows, and pricing
• get_balance — return wallet balance and granted credit

For stdio-only MCP clients, use the bridge script mcp_stdio_bridge.py:

python mcp_stdio_bridge.py https://api.inferway.ai/v1/mcp

Unauthenticated requests are mapped to the acc_guest sandbox (max 64 concurrent, no billing). Authenticated requests use the same Authorization: Bearer header as the HTTP API.

OpenAI compatibility

Inferway exposes an OpenAI-compatible /v1/chat/completions endpoint. The matrix below reflects how the gateway actually handles each parameter today — not how we wish it handled it.

Source of truth: gateway.py /v1/chat/completions handler and /v1/models capability metadata. No Pydantic request model is enforced on this endpoint; the gateway validates a small set of fields and proxies the rest.

Next steps