openweb

---

Browser automation clicks buttons, reads DOM, and burns tokens. OpenWeb calls the same APIs the website calls — when it can — and directly accesses the structured data underneath.

• Fast, cheap, and token-efficient — No screenshots, no LLM page interpretation. Structured JSON in, structured JSON out.
• Minimal effort per operation — Direct HTTP when it works, browser-backed fetch when the site requires it, and code adapters for maximal flexibility. The right transport per operation, automatically — the caller never chooses or cares.
• Predictable, typed API — Typed params, response schemas, and examples for every operation.
• Auth that just works — Cookies, JWT, CSRF, request signing, exchange chains — auto-resolved per request. You never touch tokens.
• Safe by default — Read, write, delete, and transact operations gated by permission tiers. SSRF protection on every request.
• Any site, any time — 90+ sites out of the box across social, commerce, content, travel, finance, and more. Not listed? Add it.

Quick Start

npx @openweb-org/openweb sites
npx @openweb-org/openweb wikipedia getPageSummary '{"title":"World_Wide_Web"}'

Install

CLI

npm install -g @openweb-org/openweb

openweb sites                              # list all sites
openweb wikipedia getPageSummary '{"title":"World_Wide_Web"}'

Agent Skill

Pipe this README to your agent (e.g. claude, codex) and it will self-install:

curl -fsSL https://raw.githubusercontent.com/openweb-org/openweb/main/README.md | claude

Or run the install script directly:

curl -fsSL https://raw.githubusercontent.com/openweb-org/openweb/main/install-skill.sh | bash

Works with Claude Code, Codex, OpenCode, OpenClaw, Hermes, and all agents supporting skills. After install, add to your project instructions (CLAUDE.md / AGENTS.md):

- OpenWeb: Access any website through /openweb

See skills/openweb/SKILL.md for what the skill provides.

Usage

With the agent skill

Just ask in natural language:

Use /openweb to summarize discussions on the latest AI news
/openweb Search for "laptop" across a few shopping websites

With the CLI

openweb sites                              # list all sites
openweb <site>                             # site overview + operations
openweb <site> <operation>                 # operation detail + params
openweb <site> <op> '{...}'                # execute

A managed browser auto-starts when needed. Auth-required sites may open your default browser for login.

Full CLI reference: skills/openweb/references/cli.md Troubleshooting: skills/openweb/references/troubleshooting.md

How It Works

Each site is a self-contained package:

src/sites/<site>/
  manifest.json       # identity, auth, stats
  openapi.yaml        # operations, params, response schemas
  adapters/           # JS transforms for complex sites
  examples/           # captured fixtures, used by --example and tests
  SKILL.md            # agent-facing usage notes (workflows, intent mapping)
  DOC.md              # operator notes (auth, transport, internals)
  PROGRESS.md         # site build history

The runtime reads the spec → builds the request (URL, headers, auth, CSRF) → dispatches via the right transport (HTTP, browser fetch, SSR, WebSocket) → returns structured JSON.

Auth, CSRF, and transport are declared in the spec via x-openweb extensions. Callers never configure them.

Discover

Any website can become an OpenWeb site — just ask the agent skill:

/openweb Discover and add the search function on example.com.

The agent drives the entire process — you stay in chat to make decisions when needed. All 90+ built-in sites were created this way.

Frame intents → Probe site stack → Route per family → Capture & compile (if needed) → Build package → Verify → Learn