Authentication

Durable Agent Sessions use three credential types.

Org API key

Your account key. Use it server-side for all management calls — creating agents, starting sessions, registering webhooks, managing credentials.

import { OpenComputer } from "@opencomputer/sdk";

const oc = new OpenComputer({ apiKey: process.env.OPENCOMPUTER_API_KEY! }); // org key, server-side only

const session = await oc.sessions.create({
  agent: "agt_...",
  input: "Run the tests and summarize failures.",
});

POST https://api.opencomputer.dev/v3/sessions
Authorization: Bearer $OPENCOMPUTER_API_KEY
Content-Type: application/json

{ "agent": "agt_...", "input": "Run the tests and summarize failures." }

Never ship the org API key to a browser or mobile client — it can do anything on your account. Use a client token instead.

Client tokens

A session-scoped, short-lived token with read + steer scope — safe to hand to a front-end. It can stream and steer one session and nothing else. You get one back from POST /v3/sessions, and can mint more:

const session = await oc.sessions.get("ses_...");

const clientToken = await session.mintClientToken({
  scopes: ["read", "steer"],
  ttlSeconds: 3600,
});

POST https://api.opencomputer.dev/v3/sessions/ses_.../client-tokens
Authorization: Bearer $OPENCOMPUTER_API_KEY
Content-Type: application/json

{ "scopes": ["read", "steer"], "ttl": 3600 }

Browser pattern: your backend creates the session with the org key, then hands the client_token to the browser, which uses it directly for the SSE stream and steer calls.

A client token isn’t free to hand out: each steer starts a turn, which consumes tokens on your model key for the session it’s scoped to. Mint short-lived tokens, and rely on the session’s limits (tokens/turn_seconds/turns) to cap exposure.

browser  -> your backend: create task
backend  -> OpenComputer: POST /v3/sessions (org key)
backend  -> browser: client_token
browser  -> OpenComputer: stream + steer (client_token)

For the SSE stream the browser passes the client token as ?token= (native EventSource can’t set an Authorization header); other calls use the Authorization header normally.

Model credentials

Sessions run the model either Managed — via OpenComputer, billed to your credits, with no key (the default for new orgs) — or on your own provider key (Anthropic for the claude runtime, OpenAI for codex). A bring-your-own key is stored at rest in a dedicated secrets vault (Infisical), write-only, and never returned by the API. The real key never enters a sandbox: the runtime gets an opaque placeholder, and a secret-store egress proxy swaps in the real value in flight, only on the outbound HTTPS call to the provider. A key isn’t required — Managed needs none; a session fails to start (422 no_credential) only when it resolves to neither Managed nor a usable key. For Managed, set an agent’s credential to "managed" (or rely on the org default) — nothing to add. For BYO, the quickest way to add a key is inline when you create an agent ("key": "…"). To reuse a key across agents, set an org default, or rotate/remove keys, see the dedicated Credentials page.

Getting Started

Sandboxes

Browser Sessions

Durable Agent Sessions

Reserved Capacity

CLI

Guides

TypeScript SDK

Python SDK

CLI Reference

API Reference

Resources

Org API key

Client tokens

Model credentials

​Org API key

​Client tokens

​Model credentials

Org API key

Client tokens

Model credentials