lite-locator-mcp

A semantic element finder for Playwright test agents: give it a page and a test step, get back a short list of candidate elements (with Playwright locators or MCP refs) to put in the LLM prompt instead of the whole accessibility tree. Use it as a Python library inside your own agents (LangChain/LangGraph tools included), or as an MCP proxy.

As a library (your own agents)

from lite_locator import Locator
loc = Locator()                                    # loads the bundled models once
hits = loc.find(page, "Enter standard_user in the username field")     # Playwright sync Page
hits = await loc.afind(page, step)                 # Playwright async Page
hits = loc.find(snapshot_text, step)               # Playwright MCP snapshot text you already have
hits[0]["locator"]      # page.locator('[data-test="username"]'), checked on the live page (or hits[0]["ref"])
hits.prompt             # short text for the LLM prompt
hits.tokens, hits.full_tokens, hits.fallback_full_tree   # True: send the LLM the full tree for this step

LangChain / LangGraph, with Playwright MCP through langchain-mcp-adapters (any mcp version):

from lite_locator.langchain import with_lite_locator, page_tool, LOCATOR_HINT
tools = with_lite_locator(await load_mcp_tools(session))   # adds browser_find_element, keeps every tool
tools = [page_tool(page)]                                   # or: a tool over a Playwright Python page

The base install has no MCP dependency, so it sits next to an existing mcp 1.x without changing it.

As an MCP proxy

The proxy runs Playwright MCP as a child process and exposes all of its tools unchanged, plus browser_find_element(step). That tool takes the accessibility snapshot inside the proxy, ranks the elements against the test step with small local models, and returns only a shortlist with refs. The LLM acts on a ref (browser_click target=e990) and calls browser_snapshot only when nothing in the shortlist fits.

Measured on the hard held-out test set (2,875 steps, 38 sites never seen in training), the correct element was in a shortlist of about 150–500 tokens 82–90% of the time. The full page averaged 12,229 tokens. Playwright MCP runs exactly as configured; --lean adds --snapshot-mode none, so clicks and typing no longer return a full snapshot each time.

Install

Needs Python 3.10+ and Node 18+ (for Playwright MCP itself).

pip install lite-locator-mcp          # library
pip install "lite-locator-mcp[mcp]"   # MCP proxy (needs mcp>=2.3: use its own venv if you have an older mcp)
npx playwright install chromium                           # once, if no browser is installed

Use in an MCP client

.mcp.json (Claude Code, Cursor, etc.):

{"mcpServers": {"playwright": {"command": "lite-locator-mcp", "args": ["--", "--headless", "--isolated"]}}}

Arguments after -- are passed to Playwright MCP. --lam sets the cut policy (0.002, 0.005, 0.01 or 0.02; a higher value shows fewer candidates; the default is 0.01).

Environment variables:

• PLAYWRIGHT_MCP overrides the child command (default npx -y @playwright/mcp@0.0.83).
• ELEM_RANK_LITE points to a different model bundle.

Build the wheel (from the repo)

.venv/bin/python -m elem_rank.export_onnx --out elem_rank/models-lite   # ONNX int8 models, 57 MB
.venv/bin/python -m build      # sdist + wheel in dist/

Credits and licences

• Code: MIT (see LICENSE).
• Retriever fine-tuned from BAAI/bge-small-en-v1.5 (MIT).
• Reranker fine-tuned from cross-encoder/ms-marco-MiniLM-L6-v2 (Apache-2.0).
• Training data includes Mind2Web (Deng et al., 2023, CC BY 4.0) and pages from public websites.