The archive · Developer & Business Tools · Product decision · 2026
bb-browser bets the logged-in browser is the API for AI agents; repo hits 6.2k stars
Open-source CLI + MCP server lets AI agents act through your real Chrome logins, 103 commands across 36 platforms; 6.2k stars by 2026-09-04.
bb-browser (epiral)
What the business is
bb-browser (BadBoy Browser) is an open-source CLI and MCP server that lets AI agents control your real Chrome with its login state, exposing 103 commands across 36 platforms — search, social, news, video, finance, jobs, knowledge and more — as structured JSON.
How it started
The README opens with the observation that the internet was built for browsers while AI agents are forced to use APIs — and roughly 99% of websites do not offer one, while headless browsers get detected and start logged out. So the maintainer behind bb-browser flipped the direction when the repo was created on 2026-01-31: instead of forcing websites to provide machine interfaces, let machines use the human interface directly, because 'the website thinks it's you. Because it is you.'
What happened
By the 2026-09-04 crawl the README documented 103 commands across 36 platforms — search engines, social networks, news outlets, developer sites, video, finance, jobs and knowledge — organized through a community-driven bb-sites adapter repo with one JS file per command. Adapter difficulty is tiered: cookie-based fetches in about a minute, bearer-plus-CSRF in about three, and webpack or Pinia injection in about ten. The README reports a parallel test in which 20 AI agents each reverse-engineered a website and produced a working adapter, so 'the marginal cost of adding a new website to the agent-accessible internet is approaching zero.' Distribution spans npm, an MCP server configuration for Claude Code and Cursor, and a ClawHub skill that runs inside OpenClaw's built-in browser, with the daemon speaking CDP to Chrome on localhost:19824.
No ending yet — it is still running.
Background
bb-browser (BadBoy Browser) is an open-source CLI and MCP server that gives AI agents control of your real Chrome — with its login state — instead of scraping or running headless browsers. It exposes 103 commands across 36 platforms: searching Twitter, Reddit, Zhihu or Google; pulling YouTube transcripts and Bilibili hot lists; reading Xueqiu and Eastmoney quotes; querying BOSS Zhipin jobs; summarizing Wikipedia; and dozens more, all returning structured JSON an agent can filter with jq.
The bet is stated in the README's first section: the internet was built for browsers, but agents are forced to live on APIs, and roughly 99% of websites do not offer one. bb-browser flips the direction — instead of forcing websites to provide machine interfaces, let machines use the human interface directly. Adapters run eval inside the tab, call fetch() with your cookies, or invoke the page's own modules, so the website thinks it is you because it is you — no headless browser to detect, no login to redo.
The wedge is a community adapter economy: each command is one JS file in the bb-sites repo, with three tiers of complexity from one-minute cookie fetches to ten-minute webpack injection. The README claims a test in which 20 AI agents each reverse-engineered a website and produced a working adapter, putting 'the marginal cost of adding a new website to the agent-accessible internet' near zero. Distribution runs through npm, an MCP server for Claude Code and Cursor, and an OpenClaw skill.
As of the 2026-09-04 crawl the repo stood at 6.2k stars, 603 forks and 269 commits, about seven months after creation on 2026-01-31. The material discloses no company entity, funding or revenue — the project is running live as open source, with growth itself as the observable result so far.
What has to be true
- Verifiable traction on a named platform: GitHub shows 6.2k stars, 603 forks and 269 commits at the 2026-09-04 crawl, about seven months after the repo was created on 2026-01-31.
- A crisply stated contrarian bet: while scraping stacks chase anti-bot arms races, bb-browser reuses the human interface — the README's line is that the website thinks it is you because it is you.
- A wedge with near-zero marginal cost: one JS file per command, three complexity tiers, and agents that write and submit new adapters in about ten minutes, so coverage outruns any single team.
- Ecosystem evidence inside the README: 36 platforms and 103 commands, the community bb-sites repo, npm and MCP install paths, plus an OpenClaw integration — the shape of a tool becoming infrastructure.
What can be applied
When every website already ships a working human interface, do not rebuild per-site machines: let agents reuse that session, and the product becomes the user's browser, not a scraper you maintain.
Aftermath
As of 2026-09-05 bb-browser is live and shipping in public: the repo stands at 6.2k stars, 603 forks and 269 commits, roughly seven months after creation on 2026-01-31. The README documents 103 commands across 36 platforms, a community bb-sites adapter repo, npm and MCP installation for Claude Code and Cursor, and a ClawHub skill for OpenClaw's built-in browser. The material discloses no company entity, funding, pricing or revenue — the bet that the logged-in browser is the API is still running as an open-source project.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card