Build against one governed model API.

Quickstart

OpenAI Responses example

Create a scoped API key, bind it to an application and principal, set hard limits, then use the Arc.AI base URL.

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.ARC_API_KEY,
  baseURL: "https://api.arc-ai.example/v1"
});

const response = await client.responses.create({
  model: "arc/glm-private",
  input: "Resolve this support case",
  metadata: {
    application_id: "refund-automation",
    workflow_id: "support_resolution_v3"
  }
});

Runtime APIs

Three compatible inference surfaces.

Trusted identity

Do not trust a self-reported user ID.

Production attribution comes from API-key binding, mapped JWT claims, workload identity, or signed metadata.

Legal Customer → Organization → Workspace → Application → Principal → API Key

Required:
organization_id · workspace_id · application_id
principal_id · key_id · environment

Optional:
cost_center · workflow_id · agent_id · end_user_id_hash

Request policy

Fail closed when privacy cannot be satisfied.

X-Arc-AI-Data-Policy: zdr
X-Arc-AI-Region: hk
X-Arc-AI-Log-Mode: metadata-only

A request that cannot meet the policy returns HTTP 412 with privacy_policy_unsatisfied. It never silently changes to an ineligible endpoint.

Error model

Every failure is attributable and actionable.

Claude Migration Lab

Move from a question to a version-linked rollout contract.

A migration project starts from a customer-approved Claude source manifest—not a moving model alias. It locks the exact provider snapshot, endpoint, protocol, prompt and tool hashes, source set, dataset, candidate deployments, adapters, graders, statistical plan, and budget before any comparative run starts. GLM-5.2, Kimi K3, and Tencent Hy3 deployments remain separate identities with non-transferable evidence.

Question → Sources → Lock → Harness → Run → Review → Route

Claude baseline:
  source_manifest=customer/claude-baseline@sha256:<required>

GLM-5.2:
  vendor_model=glm-5.2
  openrouter_route=z-ai/glm-5.2
  weights=zai-org/GLM-5.2@<revision required>

Kimi K3:
  vendor_model=kimi-k3
  openrouter_route=moonshotai/kimi-k3
  weights=moonshotai/Kimi-K3@<revision required>

Tencent Hy3:
  vendor_model=hy3
  provider_route=tencent/hy3
  deployment=hy3-sg-2026-07

Portability lane: preserve the Claude contract
Optimization lane: candidate-specific changes, reported separately

Decision output:
MIGRATE · PARTIAL · RETAIN_BASELINE · INCONCLUSIVE

A 220-case evaluation can support a migration decision, but production approval also requires fresh deployment and privacy evidence, customer-defined thresholds, named approvers, and explicit rollback rules.

Management surfaces

Operate the control plane through versioned resources.