Steel CLI for Browser Workflows

Short answer

Steel CLI replaces the hand-rolled scripts that slow browser automation down. It boots a named session in under a second on Steel Cloud, keeps that session permissioned, and forwards every familiar agent-browser command through one terminal contract, so you spend time finishing the workflow instead of stitching orchestration code.

If you already have Playwright or agent-browser muscle memory, the CLI adds the missing lifecycle: start, live-view, CAPTCHA solve, artifact capture, and stop, all behind steel browser <command>. Run it locally, against your own Steel Browser deployment, or in the managed cloud by flipping a single flag.

Where the CLI actually speeds things up

Bottleneck	Manual slog	Steel CLI move
Session lifecycle	Write scripts around CDP to name, reuse, and tear down contexts	`steel browser start --session run-42` and `steel browser stop` handle lifecycle for you
Proof + debugging	Capture screenshots or HLS manually	`steel browser snapshot -i`, `steel screenshot`, and `steel browser live` stream the state instantly
CAPTCHA / stealth	Bolt third-party solvers onto every run	`steel browser start --stealth` or `--session-solve-captcha` keeps solving wired in
Endpoint juggling	Rebuild configs per environment	Switch between `--local`, `--api-url`, or default cloud without changing scripts
Agent instructions	Prompt glue for every project	Install the `steel-browser` skill so agents follow the same start -> work -> stop contract

Instead of patching wrappers, treat the CLI as the workflow contract

The CLI already knows how to call Steel's session API, sanitize live-view URLs, and keep stateful runs boring. That is what makes browser work move faster. Instead of writing brittle glue around Playwright or Puppeteer, teach your agents (or your shell scripts) to call steel browser commands directly. Every command is traceable, replayable, and consistent with the docs. The steel-browser skill ships with guardrails like "always run snapshot -i before you click" so large language models stick to the safe path without ad hoc prompts.

Implementation path

Install and log in
```
npm i -g @steel-dev/cli
steel login
```
Pick the right endpoint
- Default Cloud: do nothing. Steel Cloud sessions start in under 1 second and inherit managed stealth.
- Local dev: steel dev install && steel dev start, then add --local when starting sessions.
- Self-hosted: pass --api-url https://steel.your-domain.dev/v1 so the CLI talks to your Steel Browser cluster.

Start a named session

SESSION="workflow-$(date +%s)"
steel browser start --session "$SESSION" --stealth
steel browser live --session "$SESSION"   # grab the inspect link

Drive the page with passthrough commands (identical to agent-browser syntax):

steel browser open https://example.com/checkout --session "$SESSION"
steel browser snapshot -i --session "$SESSION"
steel browser fill @e1 "Jane Doe" --session "$SESSION"
steel browser click @e5 --session "$SESSION"
steel browser wait --load networkidle --session "$SESSION"

Collect artifacts and stop cleanly

steel screenshot https://example.com/checkout --full-page
steel browser stop --session "$SESSION"

Automate the ceremony for agents
- Install the steel-browser skill: npx skills add steel-dev/cli --skill steel-browser.
- Restart your agent client so it discovers the skill manifest and enforces the same contract automatically.

Patterns that make runs faster

Signal	CLI move	Why it matters
Need parity across clouds, self-hosting, and local	Use `--api-url`, `--local`, or environment overrides	One script handles every deployment target without copy-pasted config
Agent loops keep flaking mid-action	Force named sessions plus `snapshot -i` before actions via the skill	Agents stop hallucinating selectors and you get visual evidence per step
Human review blocks progress	Use `steel browser live` during the run, then share the `connect_url` for approvals	Reviewers can jump in without stopping the automation
You keep rebuilding starter repos	`steel forge playwright --name intake-bot` and `steel run browser-use --task ...` scaffold the workflow in minutes
Credential walls slow tests	Pair CLI sessions with the Credentials API (when you move to Cloud) so secrets never live in prompts

Works for X, not yet for Y

Works when...	Still limited when...
You need deterministic browser control with evidence for every step	You expect DOM-level hooks beyond what `agent-browser` exposes; CLI mirrors that runtime
You want to run the same script against Steel Cloud, Steel Browser, or localhost	You need native mobile gestures beyond viewport emulation; use Steel mobile mode separately
You need CAPTCHAs or stealth proxies handled inside the workflow	You require bespoke proxy pools or anti-bot tricks not yet surfaced as flags
You want agent skills to follow a contract instead of bespoke prompts	You expect skills to manage secrets; pair CLI with Credentials API for storage

Next steps

Skim the Steel CLI docs for every command and flag.
Install the steel-browser skill and force your agent environment to restart so it respects the lifecycle contract.
Ship one workflow this week: run steel browser start, capture evidence with snapshot -i, stop the session, and compare how much glue code disappeared.

Humans use Chrome. Agents use Steel.