Step 01
Step 1
POST to /api/v1/messages with the standard Anthropic-shape request.
DeveloperB2Bpartial· marketing-claim · messages-api-anthropic-shape
Anthropic-shape Messages API.
Anthropic-shape /messages endpoint over our self-hosted Qwen + vLLM cluster.
The problem
Why this exists.
Developers already use the Anthropic Messages shape — tools, tool_use, tool_result, strict JSON-schema. We want them to drop their existing client onto our infrastructure without rewriting it.
Underneath is a Qwen 3.6-27B running on our own vLLM cluster — Apache 2.0 weights, no third-party API call.
How it works · 3 steps
From input to outcome.
- Step 02
Step 2
Get strict-schema tool calls validated by xgrammar; cached prefix for system, few-shots, and tool definitions.
- Step 03
Step 3
Observability and audit identical to our consumer chat surface.
Under the hood
The models and routes behind it.
Models called
API routes
- POST /api/v1/messages
Compliance posture
- Foundation model is Qwen 3.6-27B under Apache 2.0; no Claude / GPT / Gemini in the inference path.
- Every inference is metered and lands in audit-trail with the constrained-decoding path recorded.
- DPDP §6 — partner-side user data is governed by the partner's privacy notice; Mukodi is the processor.
What it's not
The Messages API surface is shipped, but the production cutover is gated on AWS GPU capacity and signed DPIA — see /models/llm_advisor for the dependency map.
Related · features & stories