Overview
Arc Documentation
Arc is an OpenAI-compatible gateway with route policy, workflow tracing, traffic controls, and observability layered on top. This section breaks the product into actual operational surfaces instead of one long reference page.
What Arc Is
Arc has three product layers: a data plane proxy in the request path, a control plane dashboard for configuration, and an operations layer for things like traces, shadow testing, smart-tier routing, and memory.
Request Path
Application
Your existing OpenAI-compatible request.
Arc Proxy
Auth, route policy, logging, memory, limits, and routing.
Provider
OpenAI, Anthropic, or another configured upstream.
Core idea
Product Model
Route
The primary traffic entrypoint. A route defines model defaults, prompts, limits, dynamic routing, memory binding, and rollout policy.
Workflow
A grouping mechanism for multi-step or agent runs. Workflows define budgets, timeouts, and max-call policy.
Trace
One workflow execution made of spans or calls. Traces let you reason about a run instead of isolated requests.
Read By Task
Get a request working
Start with the base URL swap, Arc keys, and a minimal route header.
Model customer traffic
Understand how routes act as the main policy object in Arc.
Trace agent runs
Learn how workflow headers opt requests into grouped traces.
Add continuity
Attach memory pools to routes and identify users safely.
Reduce cost carefully
Use complexity-based model routing without changing your client integration.
Inspect behavior
Use logs, traces, complexity signals, and latency breakdowns to understand traffic.
Machine-readable mirror