Browser dev tooling (vstack.browser)

Wraps the upstream chrome-devtools-mcp server so vstack can scrape agent traces from web observability dashboards, screenshot detection reports, and drive agents in the browser for evaluation.

Install

pip install 'valanistack[browser]'
npm install -g chrome-devtools-mcp     # or rely on `npx` (the default)

Chrome / Chromium must be installed on the machine. The MCP server defaults to npx chrome-devtools-mcp@latest; override with VSTACK_DEVTOOLS_MCP_COMMAND="..." if you need a different binary path.

Recognised dashboards

vstack-browser ships first-class scrape recipes for:

• LangSmith (smith.langchain.com)
• Arize Phoenix (app.phoenix.arize.com, phoenix.arize.com)
• Helicone (us.helicone.ai)
• Langfuse (app.langfuse.com)

For unrecognised URLs, the harness falls back to a "largest XHR JSON response with status < 400" heuristic and returns the most-likely candidate.

CLI

# Scrape a trace from any dashboard URL:
vstack-browser scrape "https://smith.langchain.com/o/foo/projects/bar/r/<id>"

# Take a full-page screenshot:
vstack-browser screenshot "https://example.com" --out shot.png

# List every upstream chrome-devtools-mcp tool:
vstack-browser tools

The scrape command outputs JSON with the parsed trace payload + a suggested vstack pattern target (e.g. "aar" or "lencioni" based on payload shape).

Programmatic

import asyncio
from vstack.browser import open_session, scrape_trace, screenshot_url

async def main():
    async with open_session() as session:
        scraped = await scrape_trace(
            "https://smith.langchain.com/o/.../r/<id>",
            session=session,
        )
        print(scraped.vendor, scraped.suggested_pattern)
        print(scraped.raw_payload)

asyncio.run(main())

Use cases

• Pull a real production trace from LangSmith → feed into vstack_aar (the /vstack-post-incident chain).
• Screenshot a detection report for incident channels.
• Drive a live agent UI as a black-box for evaluation.