Three API surfaces
POST /v1/responses (with a WebSocket transport for Codex-style clients), /v1/chat/completions, and /v1/messages on one upstream.
devin-2api
Devin models behind OpenAI- and Anthropic-compatible endpoints
Skip to content
Reasoning and tools round-trip
Thinking signatures replay across turns; custom tool calls and server-managed web_search work end to end.
Image, document, and video inputs
Attachments decode on all three surfaces, gated by per-model capability flags.
Detached completion cache
A client disconnect never kills a stream — an identical retry re-attaches, even across restarts.
Multi-account pool
Per-lane credentials, quota tracking, session affinity, and failover across Devin accounts.
Admin panel
Request browser, usage and quota tracking, token management, config hot reload, and one-click self-update.