Web2Actions

Turn a website into an MCP server your AI agent can use — no API required.

Web2Actions watches a website's browser traffic, reverse-engineers the site's real network endpoints, and packages them into a clean set of named tools your AI agent can call via Model Context Protocol (MCP). No public API required.

How it works

1. Capture — log into a site once; the browser traffic is recorded.
2. Generate — an LLM turns the captured traffic into a connector definition: clean named tools with inputs, outputs, and risk tags.
3. Validate — every tool is smoke-tested against the live site.
4. Serve — the validated connector is exposed as a standard MCP server.

Capture is human-driven, not automatic. We don't guess every button or click the site for you. A real browser opens on your machine; you log in and perform the actions you want turned into tools, and we record that traffic. The same philosophy applies in the hosted product — it's a popup browser you drive. The CLI just runs that browser for you.

Not supported: MFA / CAPTCHA / anti-bot challenges. If the site requires two-factor auth, a CAPTCHA, or sits behind an anti-bot challenge, Web2Actions detects it and refuses with a clear message rather than producing a broken connector.

Install

pip install -e .

Commands

Every command is web2actions <command> [options].

model — choose your LLM provider and model (persisted)

web2actions model                       # show current
web2actions model gemini/gemini-2.5-flash   # set default model (persisted)
web2actions model --provider openrouter # set default provider
web2actions model --alias sonnet=anthropic/claude-sonnet-4  # alias
web2actions model --aliases             # list aliases

analyze — discover a site's API surface (agent harness)

Runs the agent over captured traffic and lists the meaningful API endpoints (using your configured LLM — any provider via litellm). It reverse-engineers almost any backend — a custom REST API, GraphQL, SDK/RPC protocols, form-based server actions, or a backend on a completely different domain (like Supabase) — by analyzing the real traffic rather than assuming a shape:

web2actions capture https://app.example.com          # produces traffic.json
web2actions analyze traffic.json                      # shows the API surface (asks clarifying questions)
web2actions analyze traffic.json --url https://app.example.com -o connector.json
web2actions analyze traffic.json --assume             # no prompts — use recommended defaults

You can pick a model with --model, or it uses your saved web2actions model choice. During analysis it may ask a few clarifying questions as numbered menus (like Claude Code) — use the number or just press Enter for the recommended option; --assume skips them for automation. With -o, analyze writes a connector-definition JSON (validated against our schema) that you can then validate and serve as MCP.

capture — record a site's traffic

Opens a browser to the URL, records the network traffic you generate as you log in and click, then writes a JSON dump.

# Open the site; log in and click around manually; press Enter when done.
web2actions capture https://app.example.com

# Filter out analytics/static noise and save to a chosen file.
web2actions capture https://app.example.com --filter -o dump.json

# Automate a simple form login (fill + submit), then continue capturing.
web2actions capture https://app.example.com --login \
  --username you@example.com --password "..." \
  --username-selector "#username" --password-selector "#password" \
  --submit-selector "#submit-btn" --success-indicator "#dashboard"

Options: --login, --username, --password, --username-selector, --password-selector, --submit-selector, --success-indicator, --filter, --output/-o.

generate — turn captured traffic into a connector definition

web2actions generate dump.json --model claude-sonnet-4 -o connector.json

Options: --model (or WEB2ACTIONS_MODEL env), --output/-o.

validate — check a connector against the schema

web2actions validate connector.json
# -> VALID

serve — expose a connector as an MCP server (stdio)

web2actions serve connector.json

Wires the connector to your AI agent's MCP client over stdio.

End-to-end example

web2actions capture https://example.com --filter -o dump.json
web2actions analyze dump.json --assume -o connector.json   # agent discovers API → connector
web2actions validate connector.json
web2actions serve connector.json

Works for almost any backend (a custom REST API, GraphQL, a Supabase-backed app, form server-actions, a backend on a different domain) — capture records all hosts, the agent reverse-engineers the real API, and the result is served as MCP. Sites behind MFA/CAPTCHA/anti-bot are refused with a clear message.

Start from an existing example (no capture needed)

web2actions validate connector-spec/examples/simple-crm.json
web2actions serve connector-spec/examples/simple-crm.json

Bringing your own LLM (BYOK)

Web2Actions uses litellm, which connects to 100+ providers through one OpenAI-compatible interface. You pick a provider and a model with simple commands — no code changes, and the choice is remembered for future runs.

Supported providers

litellm supports 100+ providers, including OpenAI, Anthropic, Google Gemini, OpenRouter, DeepSeek, Groq, xAI, Mistral, OpenAI-compatible local servers (Ollama / llama.cpp / vLLM), Bedrock, Azure, and more.

Each provider reads its standard API-key environment variable (for example OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY, OPENROUTER_API_KEY, DEEPSEEK_API_KEY, GROQ_API_KEY, XAI_API_KEY) or, for local servers, needs no key at all.

Set the default model with a simple command

# Pick a provider + model once; it is remembered (saved to ~/.web2actions/config.toml)
web2actions model openai/gpt-4o-mini
web2actions model anthropic/claude-sonnet-4
web2actions model gemini/gemini-2.5-flash
web2actions model openrouter/anthropic/claude-sonnet-4
web2actions model deepseek/deepseek-chat

# Show the current model
web2actions model

# Set a default provider
web2actions model --provider openrouter

# Add a short alias, then use it anywhere
web2actions model --alias sonnet=anthropic/claude-sonnet-4
web2actions generate dump.json --model sonnet

# List your aliases
web2actions model --aliases

Once set, analyze uses that model automatically:

web2actions model gemini/gemini-2.5-flash
export GEMINI_API_KEY=...            # only needed the first time, per provider
web2actions analyze traffic.json --assume   # uses gemini-2.5-flash

Override per run

Pass --model to use a different provider/model for a single command, or set the WEB2ACTIONS_MODEL env var. Precedence: --model > WEB2ACTIONS_MODEL > saved config.

web2actions generate dump.json --model openai/gpt-4o-mini

If no key is set

generate fails gracefully and tells you which key to set — it never crashes with a raw stack trace.

Repository layout

• connector-spec/ — the connector definition schema + validator
• capture/ — browser session + traffic recording + noise filtering
• generate/ — LLM extraction (cheap path) + sandboxed fallback
• validate/ — live tool smoke tests + risk tagging
• mcp-runtime/ — the shared MCP server that serves any connector
• agent/ — the agent harness (vendored CLI-Anything-Web) + provider-agnostic LLM backend
• cli/ — the web2actions command-line wrapper

License

Apache-2.0. Built on top of the open-source CLI-Anything / CLI-Anything-Web projects.