unbrowse-ai/unbrowse

Unbrowse - Reverse engineer any website so that your agents can access its APIs directly

Unbrowse – Fast, cheap agentic browsing

What it is – An open‑source TypeScript client (@unbrowse/sdk) and CLI that let AI agents talk to a website’s first‑party HTTP API instead of driving a full browser. By re‑using the site’s own backend calls, Unbrowse claims 50‑100× speed‑ups and up to 95 % cost reduction.

How it works

  1. Discover – The tool looks at your private capabilities and a public registry of pre‑compiled site tools.
  2. Run – If a matching capability exists, Unbrowse replays the site’s native HTTP requests in a single call. The result is verified before being returned.
  3. Learn – When no capability matches, the system records a browser session (or you upload two HAR files). Those recordings are compiled into a new one‑call capability for future use.

Key components

  • CLI (npx unbrowse …) – login, discover, run tasks, scrape pages, index sites, learn from HAR recordings, view saved logins, query the public registry.
  • SDK (@unbrowse/sdk) – programmatic access from TypeScript/JavaScript; example shows creating a Unbrowse instance, running a task, and awaiting a verified result.
  • MCP (Managed Cloud Proxy) – hosted service (https://unbrowse.ai/mcp) that agents can connect to via HTTP or a local stdio proxy (unbrowse mcp).
  • Plugins – ready‑made bundles for Claude, Codex, Grok, OpenClaw, Hermes, elizaOS, and CLIs for ChatGPT, Perplexity, Gemini, etc., making Unbrowse act as the agent’s browser.

Typical workflow

# add the skill to an agent
npx skills add unbrowse-ai/unbrowse --skill unbrowse
# register the MCP endpoint for Claude
claude mcp add --transport http unbrowse https://unbrowse.ai/mcp
# sign‑in (OAuth opens in browser)
npx unbrowse login
# run a task
npx unbrowse run 'top stories on hacker news'

If the site isn’t yet in the registry, you can record it:

unbrowse learn a.har b.har   # turn two HAR recordings into a reusable capability

Pricing & quotas – The CLI can show usage (unbrowse usage). Pricing details are in the linked docs; the service is hosted, while the client and SDK are MIT‑licensed open source.

Development – Uses bun for package management, testing, type‑checking, and building. bun test packages runs SDK and CLI tests against a local mock API.

Documentation links – Install guide, CLI reference, MCP usage, SDK reference, troubleshooting, white‑paper “Internal APIs Are All You Need” (arXiv:2604.00694), and open‑source policy.

Bottom line – Unbrowse provides a practical way for LLM‑based agents to interact with web services quickly and cheaply by leveraging the sites’ own APIs, with a CLI/SDK for developers and ready‑made plugins for major AI assistants.

Related

  • Project
    Aas-ee/open-webSearch

    Open‑WebSearch is a Node‑JS service that lets AI agents perform live web searches and fetch article content without any API keys. It offers an MCP‑compatible server, a CLI, and a long‑running HTTP daemon, aggregates results from many public search engines, supports proxy use, and can optionally use Playwright‑driven headless browsers for sites that need JavaScript rendering.

  • Project
    browser-use/browser-use

    Browser Use is an open‑source Python/TypeScript library that gives LLM agents full control over a real or cloud‑hosted web browser. It provides 31 browser actions as tools, supports custom functions, works with OpenAI, Anthropic, Google models or its own BU2 model, and offers three usage modes: a fully hosted cloud API, a CLI for existing agents, and a library you embed in your code. Ideal for automating web tasks, building AI assistants that browse, or researching LLM‑driven web navigation.

  • Project
    anomalyco/browser-control

    Browser Control is an open‑source Node.js driver that lets AI agents or CLI users run Playwright code inside their own Chromium browser (with all logged‑in sessions and extensions). It works via a local relay and a browser extension, supports persistent sessions, inspection helpers, human hand‑offs, video/ HAR recording, and a TypeScript Effect client for secret‑safe API calls. Intended for trusted local use only.

  • Project
    platonai/Browser4

    Browser4 is an AI‑native browser automation engine that lets LLM agents (or humans) drive real browsers via a fast Rust CLI, extract data with zero‑token X‑SQL/CSS selectors, and run large‑scale crawls or ML‑based spreadsheet generation (WebMiner). It combines CDP‑native performance, swarm scaling, and hybrid LLM/ML intelligence, and is distributed as an npm CLI with a Java backend (Maven Central).

  • Project
    chainbase-labs/Agentkey

    AgentKey is a genuine open‑source tool that adds one‑click, subscription‑based internet access to many LLM agents. It installs a skill and a tiny MCP client, authenticates via a browser, and then lets the agent fetch tweets, Reddit threads, YouTube transcripts, on‑chain data, stock quotes, product listings, etc., without the user handling individual API keys.