@false00/pi-steel

Fork of @steel-experiments/pi-steel by nibzard and the Steel Experiments team. All credit for the original work belongs to them.

	Upstream	This fork
npm	@steel-experiments/pi-steel	@false00/pi-steel
GitHub	github.com/steel-dev/pi-steel	github.com/false00/pi-steel
Pi registry	@steel-experiments/pi-steel	—

Steel browser automation tools for the Pi coding agent.

This package publishes the Steel extension as a reusable Pi package so it can be installed directly into Pi or consumed by other runtimes such as Takopi-based wrappers.

Changes from upstream

• .env file support — ~/.config/steel/.env is read as the highest-priority config source for api_key and base_url. Auto-created on first run with discovered values so you don't have to set env vars every time.
• Tool output paths — steel_screenshot, steel_computer, and steel_scrape all append \nPath: {absolute-path} to their visible text output so the LLM can read the artifact directly. Upstream hid these paths in details metadata that the agent never sees.
• steel_scrape saves full content to disk — the untruncated scraped content is written to ~/.cache/.steel-browser/scrapes/ before the inline output is trimmed. When the response shows [truncated N chars], the LLM can read the file for the complete content.
• Artifact directories moved — screenshots and PDFs are saved under ~/.cache/.steel-browser/ instead of $CWD/.artifacts/, keeping build artifacts out of project directories.

Quick start

pi install npm:@false00/pi-steel

Then just ask Pi to browse:

> Go to hacker news and find the top story

Pi will use steel_navigate to open the page, steel_screenshot/steel_scrape to read the content, and return what it finds. All session management happens automatically.

Tools

Navigation

Tool	Description
steel_navigate	Open a URL with automatic scheme normalization and retry logic
steel_go_back	Navigate back in browser history
steel_get_url	Read the current page URL
steel_get_title	Read the current page title

Content extraction

Tool	Description
steel_scrape	Extract page content as text, markdown, or html (full content saved to disk with path in output)
steel_screenshot	Capture a screenshot artifact (path in output)
steel_pdf	Generate a PDF artifact
steel_extract	Extract structured data using a JSON schema

Interaction

Tool	Description
steel_click	Click an element with captcha recovery
steel_type	Type text into a field
steel_fill_form	Fill multiple form fields at once
steel_scroll	Scroll the page or a nested container
steel_find_elements	Find interactive elements by selector
steel_wait	Wait for an element to appear
steel_computer	Low-level computer action with screenshot (path in output)

Session management

Tool	Description
steel_pin_session	Keep the browser session alive across prompts
steel_release_session	Close the browser and reset to default session mode

steel_scrape defaults to text. Ask for markdown when headings, lists, and links matter. Ask for html only when raw DOM markup is actually needed.

Every steel_scrape call saves the full content to ~/.cache/.steel-browser/scrapes/ and includes the file path in the tool output. If the inline response is truncated, the LLM can read the file to get the complete content.

steel_scroll can scroll the page or a nested container. For apps like Google Maps, pass a selector for the results pane instead of relying on window scrolling.

Install

Install into Pi as a package:

pi install npm:@false00/pi-steel

Or load it for a single run:

pi -e npm:@false00/pi-steel

For local development from this repo:

pi -e .

Session modes

Steel sessions have a lifecycle tied to how Pi uses them. The default works for most cases, but you can tune it:

Mode	Behavior
agent (default)	One session per Pi prompt, closed after agent_end
session	Session stays alive until Pi switches or shuts down
turn	Session closed after each Pi turn — aggressive, can break multi-step workflows

Set the mode via environment variable:

STEEL_SESSION_MODE=session pi -e npm:@false00/pi-steel

You can also change session persistence at runtime with steel_pin_session and steel_release_session.

Configuration

Required

• Node.js 20+
• A Pi runtime that supports extensions
• Steel authentication via one of:
  • .env file in ~/.config/steel/ (highest priority), or
  • STEEL_API_KEY, or
  • steel login config in ~/.config/steel/config.json

You can also create a ~/.config/steel/.env file to set your API key and base URL:

api_key=your-steel-api-key
base_url=https://your-selfhosted-steel-instance.com

Values in this file take precedence over all other configuration sources, including environment variables and constructor parameters.

Environment variables

Connection

Variable	Purpose
~/.config/steel/.env	api_key and base_url values (highest priority)
STEEL_BASE_URL	Steel API base URL
STEEL_BROWSER_API_URL	Browser API endpoint
STEEL_LOCAL_API_URL	Local Steel instance URL
STEEL_API_URL	Alternative API URL
STEEL_CONFIG_DIR	Custom config directory

Session

Variable	Purpose
STEEL_SESSION_MODE	Lifecycle mode: agent, session, or turn
STEEL_SESSION_TIMEOUT_MS	Session timeout
STEEL_SESSION_HEADLESS	Run browser headless
STEEL_SESSION_REGION	Browser region
STEEL_SESSION_PROFILE_ID	Persistent browser profile
STEEL_SESSION_PERSIST_PROFILE	Save profile changes
STEEL_SESSION_CREDENTIALS	Session credentials
STEEL_SESSION_NAMESPACE	Session namespace

Proxy

Variable	Purpose
STEEL_USE_PROXY	Enable proxy
STEEL_PROXY_URL	Proxy URL

Captcha

Variable	Purpose
STEEL_SOLVE_CAPTCHA	Enable captcha solving
STEEL_CAPTCHA_MAX_RETRIES	Max captcha retry attempts
STEEL_CAPTCHA_WAIT_MS	Captcha solve wait time
STEEL_CAPTCHA_POLL_INTERVAL_MS	Captcha poll interval

Tools

Variable	Purpose
STEEL_TOOL_TIMEOUT_MS	Default tool timeout
STEEL_NAVIGATE_RETRY_COUNT	Navigation retry attempts

pi-steel reads Steel CLI config for auth and local API resolution, and it normalizes CLI-style API URLs such as http://localhost:3000/v1 to the SDK-compatible base URL form.

Development

npm install
npm run build   # requires tsconfig.json — only if you have TS source
npm test        # requires tsconfig.json — may not work in this fork

Publishing

This fork's dist/ is the source of truth (no TS source files). The npm test and prepublishOnly scripts require a tsconfig.json that doesn't exist in this repo, so publish with --ignore-scripts.

# 1. Bump version and create git tag
npm version patch   # or minor / major

# 2. Verify package contents
npm pack --dry-run

# 3. Publish (skips broken prepublishOnly script)
npm publish --ignore-scripts

If 2FA is enabled, npm will prompt you to authenticate in the browser before publishing.

The package manifest in package.json exposes the compiled extension entrypoint via pi.extensions, which lets Pi load the package root directly after install.

License

MIT — same as upstream. Original work copyright (c) 2026 Steel (nibzard & Steel Experiments contributors).