Web Capability Compiler (WCC)

Compile existing websites into reusable, typed capabilities for MCP agents.

WCC is not another browser-use agent. It is a compilation, binding, and caching layer that sits between an existing website and an external AI agent.

Browser agents typically re-interpret low-level UI elements on every task, which costs tokens, reduces repeatability, and ties workflows to transient DOM detail. WCC inspects a page once, compiles the observable controls into a small set of semantic, typed, low-risk capability templates, binds compatible templates to the current session, and lets an MCP client execute them through a stable interface. On repeated use it tries to bind a persisted template before recompiling, so the agent can skip low-level DOM interpretation entirely.

The headline metric is token reduction on repeated tasks, not task success rate.

Status

MVP v0.1 is under construction against spec v0.2.1. Phases 0–6 of 7 are complete, and it runs as an MCP server today. The loop closes: WCC reads a live document into a snapshot, builds an Agent Page IR, compiles typed search, filter, and navigation capabilities with evidence and a deterministic risk classification, persists them as reusable templates, binds a stored template on a later session without recompiling, and executes one — validating arguments against its schema, re-resolving every target, re-checking risk against the page in front of it, running the plan one allowlisted action at a time, and verifying afterwards that the intended effect actually happened.

All of that is reachable from any MCP client over stdio, through six stable tools. An optional LLM pass can improve the names and descriptions, and is off by default — nothing in the loop above depends on it. What remains is Phase 7, the token-reduction benchmark.

Three properties are worth stating plainly, because they are what the design is for: only low-risk capabilities ever execute; an execution is reported as successful only when its effect was observed; and no selector, screenshot, or coordinate crosses the MCP boundary in either direction — a test asserts the last one by searching every tool response for the fixtures' own selectors.

See docs/progress.md for the phase log, which records the reasoning behind each decision and the defects found along the way.

Quickstart

No clone required. uv and Python 3.12+ are the only prerequisites.

# Install the Chromium build the driver uses (once)
uvx --from git+https://github.com/Maaa2005/web-capability-compiler.git wcc install-browser

# Check the machine before pointing a client at it
uvx --from git+https://github.com/Maaa2005/web-capability-compiler.git wcc doctor

# Run the MCP server over stdio
uvx --from git+https://github.com/Maaa2005/web-capability-compiler.git wcc serve

Connecting an MCP client

Claude Code:

claude mcp add wcc -- uvx --from git+https://github.com/Maaa2005/web-capability-compiler.git wcc serve

Claude Desktop (claude_desktop_config.json) or any client that takes a command:

{
  "mcpServers": {
    "wcc": {
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/Maaa2005/web-capability-compiler.git",
        "wcc",
        "serve"
      ]
    }
  }
}

Then ask the agent to open a page and inspect it. It will get back a list of typed capabilities rather than a DOM.

The six tools

Tool	What it does
open_session	Open a browser session on a URL, subject to the URL policy
inspect_page	Compile the page — or rebind a stored template — and report capabilities, entities, and refusals
list_capabilities	Filter what the inspection found by category and maximum risk
execute_capability	Run one capability and return the complete verified result
get_execution_result	Retrieve a stored result, only for the session that produced it
close_session	Close the session; idempotent

Generated capabilities are returned as typed data, not registered as MCP tools — so the client's tool list never changes as the agent browses.

Configuration

Set by whoever launches the server, never by an agent over the wire.

Variable	Effect
WCC_DATA_DIR	Where templates and history are stored. Absolute paths only; defaults to the OS per-user data directory.
WCC_ALLOW_PRIVATE_NETWORK	Set to 1 to allow loopback and private addresses. Development only — it switches off the SSRF boundary of specification 10.2.
WCC_ENRICHMENT	Set to 1 to let an LLM improve capability names and descriptions. Off by default; needs pip install 'wcc[llm]' and ANTHROPIC_API_KEY.
WCC_ENRICHMENT_MODEL	Which model does that naming. Defaults to claude-opus-5.

Optional semantic enrichment

Off unless you turn it on, and the compiler is the product either way — enrichment only rewrites names and descriptions. What it is given is a reduced structured summary: the compiled capability names, categories, risk levels, and parameter names. It is never given the DOM, a selector, a page's text, or a form value. What it gets back is validated before use and cannot change a locator, an execution plan, or lower a risk classification; a risk hint is only ever applied upward. If the provider is unreachable or answers with something invalid, the deterministic result is used unchanged and the reason is logged.

Responses are cached by page signature, so one page class costs one call.

Operating it

uv run wcc diagnostics                        # sanitized counts and environment checks
uv run wcc export --session <id> --output caps.json
uv run wcc import caps.json                   # checked on the way in, re-verified at bind time
uv run wcc clear-data                         # remove all local WCC data

An imported capability is stored, never trusted: it re-resolves its targets and is re-classified for risk against the live page before it can run, exactly as a locally compiled one is.

Development

uv sync
uv run wcc --help
./scripts/test.sh

Scope of MVP v0.1

Supported capability categories are search, filter, and navigate. Only low risk capabilities are executable. Purchases, account changes, message sending, and any state-changing form submission are out of scope, as are CAPTCHA bypass, anti-bot evasion, and credential storage.

The full scope, data model, risk model, and benchmark methodology live in docs/spec/WCC_AGENT_IMPLEMENTATION_SPEC_v0.2.1.md, which is the source of truth for implementation.

License

MIT