About Steel
Steel runs browsers in the cloud and gives you an API to create, control, watch and release them. Each browser is a session. Your code or agent decides what the browser does. Steel runs the machine, browser, connection and live view, and keeps each run's records.
The parts
your code or agent
|
| HTTPS to api.steel.dev: create, Commands, Quick actions, release
| WebSocket to connect.steel.dev: CDP for Playwright and Puppeteer
v
session: one browser in its own virtual machine -----> the web
|
+-- live view: the session viewer, while the session runs
+-- after it ends: recording, logs, agent traces, signals, files| Part | What it is |
|---|---|
| API | The HTTP API at https://. Every request carries your API key in the steel-api-key header. The Steel SDK, the Steel CLI and the MCP server all call it. |
| Session | One browser that Steel runs for you. It has an id, a status and a timeout. |
| Connection | The way your code drives the browser: over CDP with Playwright or Puppeteer, or over HTTP with Commands or Computer actions. Control interfaces compares them. |
| Live view | The browser in real time: in the Steel dashboard at sessionViewerUrl, or by itself at debugUrl, which you can embed in your app. Live view shows both. |
| Recording | A video of the session that you can replay after it ends. |
What happens when you create a session
You send POST /v1/sessions, or call sessions.create in the SDK. The request can set options such as the timeout, a proxy or a profile. Steel starts a browser with those options and returns the session.
The response holds what you need to use the session:
id: the session's ID. Steel generates it.status:liveonce the browser runs.websocketUrl: the CDP URL for Playwright or Puppeteer.sessionViewerUrl: the session viewer in the Steel dashboard.
From then on, the session runs until you release it or a timeout ends it. Session lifecycle explains how a session ends.
Each live session counts toward the number of concurrent sessions that your plan allows. Limits has the number for each plan.
Sessions are isolated
Each session is one browser in its own virtual machine. Two sessions share no cookies or storage. To start a session with an earlier session's state, load a profile or a session context. Browser state compares the ways to do it.
What a session leaves behind
When a session ends, the browser stops, but Steel keeps the run's records:
| Record | What it holds |
|---|---|
| Recording | The session's screen, to replay. Sessions record by default. Set recording to false to turn it off. |
| Logs | The session's browser events. |
| Agent traces | A timeline of the clicks, typing, scrolls and navigations in the session. |
| Signals | Labels for events that Steel saw, such as a CAPTCHA or a bot block. Steel adds them a few seconds after the session ends. |
| Files | The files that you uploaded to the session and the files that it downloaded. |
| Session cost | The session's final cost, once it is released. |
If the session saved a profile, the profile also stays, for the next session to load.
Steel keeps a session's records for your data retention period. Data retention has the period for each plan.