Awesome Real-Browser Agents

A curated collection of practical AI agent workflows that use a real browser session — not just search snippets, APIs, or scraped HTML.

Real-browser agents are useful when the work requires logged-in websites, dynamic pages, search result inspection, social platforms, ecommerce marketplaces, competitor pages, SaaS dashboards, and human-controlled cookies/permissions.

This repo collects copy-paste prompts, source maps, workflow templates, and examples for agents that need to operate across the real web.

The core idea: give the agent scoped browser access, make it collect evidence first, then require approval and a receipt before risky actions.

Why real-browser agents?

Many agent workflows fail because they start from a blank chat prompt. A better workflow starts with evidence:

1. Search the real web.
2. Open the source pages.
3. Use the website like a human operator.
4. Save URLs and excerpts.
5. Structure the findings.
6. Only then write the report, update the CRM, draft the reply, or recommend the next action.

Real-browser access matters because many useful business signals are not available as clean APIs: Reddit threads, X/LinkedIn/Product Hunt pages, Amazon/Taobao/Tmall/Shopify/Etsy pages, G2/Capterra reviews, competitor pricing pages, docs, changelogs, CMS dashboards, support inboxes, and internal SaaS tools.

BrowserMan

BrowserMan connects AI agents to your real Chrome browser.

It is designed for delegated browser access:

• real logged-in sessions,
• agents can run anywhere,
• cookies stay local in your browser,
• scoped permissions,
• audit trails,
• one-click revoke.

Safety model: scope, gates, receipts

Real-browser agents are not only reading pixels. When they use logged-in apps, they are operating inside delegated authority. Treat the browser session like a capability.

Each workflow in this repo should make three boundaries explicit:

1. Scope — which browser session, websites, tabs, records, queues, or drafts the agent may inspect.
2. Gates — which actions require human approval before execution, such as send, publish, refund, export, merge, delete, update-live, or trigger-outreach.
3. Receipts — what the agent hands back after the run: source URLs, changed fields, proposed actions, approvals requested, skipped risky actions, and state left behind.

The goal is not an invisible agent that clicks faster. The goal is an auditable workflow that a human can review, approve, revoke, and resume.

You can use BrowserMan from agents through MCP, HTTP API, optimized site scripts, or higher-level orchestrators such as OpenClaw.

Useful links:

• BrowserMan: https://browserman.run
• BrowserMan Blog: https://blog.browserman.run
• OpenClaw: https://openclaw.ai
• ClawMama hosted OpenClaw: https://clawmama.run

Workflows

Workflow	What it does	Best for
Market Research	Reads real customer, competitor, review, and marketplace sources before generating a brief	founders, PMMs, researchers
Competitor Research	Builds battlecards from pricing pages, changelogs, reviews, docs, and community mentions	SaaS, GTM, sales enablement
Daily Brand Radar	Checks Google, social profiles, competitor pages, and fresh mentions every morning	operators, marketers, founders
Ecommerce Product Research	Collects product titles, prices, claims, reviews, and marketplace trends	ecommerce, DTC, marketplace teams
Social Listening	Finds repeated questions, complaints, and language across social/community surfaces	marketing, support, product
Logged-in SaaS Ops	Lets an agent work inside tools such as CRM, CMS, support inbox, analytics, and admin dashboards	ops, support, growth
Support Inbox Approval Queue	Triages support tickets, drafts replies, and queues refunds/sends/account changes for human approval with receipts	support, ops, CX
Lead Research with a Real Browser	Researches prospects across approved logged-in tools and public sources, then hands off sourced lead briefs before outreach	sales, growth, founders
CMS Draft / Publish Gate	Prepares CMS drafts, validates metadata, and queues publish/update/delete actions for approval with source-context receipts	marketing, content, ops
CRM Contact Update Receipt	Inspects CRM records, proposes field-level updates with evidence, and queues edits/exports/outreach for approval	sales, ops, CRM

Quick start

1. Pick a narrow job

Bad:

Analyze my market.

Better:

Research AI support tools for B2B SaaS. Collect Reddit alternative threads, G2 complaints, competitor pricing pages, changelogs, and launch posts. Save URLs and excerpts before writing conclusions.

2. Connect a real browser

Install BrowserMan from https://browserman.run and connect your Chrome browser.

npm install -g browserman-cli
browserman setup
browserman doctor
browserman browser ping --json

Or with npx:

npx -y browserman-cli setup
npx -y browserman-cli doctor
npx -y browserman-cli browser ping --json

3. Give your agent the workflow prompt

Open one of the workflow folders and copy prompt.md.

4. Require evidence before conclusions

Every useful real-browser agent should save URLs, titles, excerpts, timestamps, source types, confidence levels, and structured rows. If the agent cannot point to the source, it should not make the claim.

Example BrowserMan Google search

browserman script run \
  --site google.com \
  --action search \
  --text "AI agent competitor research G2 Reddit pricing pages" \
  --wait \
  --json

See more examples in examples/.

Principles

1. Evidence first, generation second

Do not let the model invent the market from memory. Make it collect sources first.

2. URLs or it did not happen

Every claim should have a source URL or be clearly marked as inference.

3. Use scoped browser access

Do not paste passwords or cookies into prompts. Use delegated access and revoke it when done.

4. Gate risky actions

Reading, drafting, and collecting evidence can often run quickly. Sending, publishing, refunding, exporting, deleting, merging, or updating live records should be explicit approval boundaries.

5. Produce a receipt

A useful agent run leaves a handoff artifact: what it inspected, what it changed or proposed, what still needs approval, and what state the browser/app was left in.

6. Keep workflows small

Start with one brand, one market, three competitors, or ten source pages. Expand once the loop is reliable.

7. Separate collection from analysis

A good workflow has at least two phases: collect raw evidence, then synthesize patterns and recommend action.

Contributing

Pull requests are welcome. Good additions include real-browser workflow prompts, source maps, BrowserMan/OpenClaw examples, operator playbooks, screenshots or traces with private data removed, and notes on what failed and how to make it reliable.

Please keep examples practical and evidence-based. Avoid generic “AI will automate everything” content.

Related reading

• Browser agents should be auditable, not undetectable
• How to Build an AI Agent for Market Research with a Real Browser
• Build a Daily Brand Radar Agent with OpenClaw and BrowserMan
• 用 OpenClaw 和 BrowserMan 搭一个每天 8 点汇报的品牌雷达 Agent