rzn-browser

name: rzn-browser description: Browser automation for data extraction, form filling, search, posting, messaging, and signed-in web workflows. Use direct CLI commands to drive the browser through the native host bridge, or run premade workflows like Google Search, Reddit post/comment flows, PubMed search/extract, ChatGPT/Claude chats, X threads, Instagram assets, Amazon/G2/Capterra reviews, and App Store lookup.

RZN Browser

Use direct CLI commands to automate the user's browser through the native host bridge. This skill is both a use skill and a workflow-build skill:

• Use mode: run existing workflows or llm-auto to complete browser tasks.
• Build mode: create, edit, validate, or promote workflow JSON when a task should become repeatable.

Prefer direct CLI commands. Do not write Python helper scripts or wrapper scripts for normal use.

Use This When

• The user asks to automate, inspect, extract data, fill forms, search, post, message, or operate a signed-in website.
• A premade workflow might already cover the task.
• The task is fuzzy enough for llm-auto, but still belongs in the browser.
• A successful fuzzy run should become a reusable workflow.
• The user asks to build, debug, validate, or contribute a workflow.

Do Not Use This When

• A normal answer, API call, file edit, or local code change solves the task without browser automation.
• The user needs a test suite, remote browser infrastructure, or large-scale scraping.
• The browser/native-host bridge is unavailable.
• The task would post, send, buy, like, submit, delete, or otherwise mutate state without explicit approval or a draft/review mode.

Good Premade Workflow Examples

• Search/research: Google Search, Google Scholar, Google Maps, PubMed search/extract, arXiv, ScienceDirect.
• Posting and communities: Reddit post/comment/message flows, Hacker News submit/comment/reply, X post/thread/DM flows.
• AI apps: ChatGPT send/export/download flows, Claude send/recent/export flows.
• Shopping and reviews: Amazon product/search, G2 reviews, Capterra reviews, Etsy listings.
• App and social data: App Store lookup/search, Instagram profile/post asset extraction.

First Move

1. Inspect the catalog before improvising:

rzn-browser list
rzn-browser list <system>
rzn-browser list <system> <workflow>

2. If a listed workflow fits, run it:

rzn-browser run <system> <workflow> --param key="value"

3. If the task is ambiguous or no workflow fits, use llm-auto:

rzn-browser llm-auto "Do the browser task in plain language" --max-steps 20

4. If the task is to create or improve a durable workflow, read references/workflow-authoring.md.

Read references/cli-cheatsheet.md for exact command forms and flags.

Choice Rule

Runtime Contract

• Use the user's existing browser session with the extension and native host.
• Do not launch isolated profiles, temporary browser apps, or a fresh browser unless the user explicitly asks.
• If native host or extension wiring is broken, stop and tell the user what to reload instead of silently switching tools.
• Use rzn-browser on PATH. In a source checkout, ./target/release/rzn-browser is acceptable only when the installed binary is missing or the task needs local uninstalled changes.
• llm-auto needs real provider env vars (OPENAI_* or GEMINI_*) unless using LLM_PROVIDER=dummy for local deterministic smoke checks.
• For machine-readable output, prefer --json where the command supports it.

Read references/runtime-troubleshooting.md if a command cannot connect to Chrome or the native host.

Workflow Flow

User task
  |
  v
Check catalog: rzn-browser list [system] [workflow]
  |
  +-- matching workflow --> rzn-browser run ... --param ...
  |
  +-- no match / fuzzy --> rzn-browser llm-auto "..."
                              |
                              +-- worth reusing --> save workflow, validate, rerun

Working With Workflows

Use workflow ids in the form <system> <workflow> or system/workflow. Prefer installed ids over repo-relative JSON paths once a workflow is part of the catalog.

Common inspection and validation:

rzn-browser workflow show <system> <workflow> --json
rzn-browser workflow validate <system> <workflow>
rzn-browser workflow dirs

If editing or creating workflow JSON, read references/workflow-authoring.md first. Existing workflows must keep site-specific details in workflow data/docs, not in shared engine code.

llm-auto Defaults

Use llm-auto for goals, not for known fixed workflows:

rzn-browser llm-auto "Search Google for browser automation tools and extract the first page" --json

Useful flags:

• --url <url> to start somewhere deterministic.
• --context "..." to provide constraints or known state.
• --constraint "..." to make rules explicit.
• --max-steps <n> to bound execution.
• --prefer-cached false when you do not want cached workflows tried first.
• --save-workflow true --name <name> when a successful run should become reusable.
• --pure-llm only when deterministic macros are getting in the way.

Reporting Results

Report:

• command used
• whether it used a workflow or llm-auto
• important output rows or extracted fields
• any browser-side action left for user review
• connection/runtime failures with the exact next manual step

Do not paste huge raw JSON unless the user asks. Summarize the high-signal fields and mention how to rerun the command.