Chrome DevTools MCP through mcp2cli

A skill for agents that need to drive Chrome from a terminal, including agents without native MCP integration.

It wraps Chrome DevTools MCP through mcp2cli, giving shell-first agents a repeatable way to start a browser session, navigate, inspect page state, interact with forms, read console/network evidence, and clean up.

It is meant for browser smoke tests, app-driving checks, and focused debugging from command-line-capable agents. Use direct MCP when your agent has native MCP support and you want the full typed tool surface inside the agent runtime.

Field note

This skill came out of a small Copilot CLI experiment comparing direct Chrome DevTools MCP with a CLI-wrapper path over the same 9-step browser smoke test (3 runs per variant).

| Path | Fresh context | Completed-run message growth | Runtime notes | |---|---:|---:|---| | CLI skill through mcp2cli | 19k total, MCP Tools 155 | about 18-21k | 9/9 passes; slower in one run, steadier context growth | | Direct Chrome DevTools MCP | 24k total, MCP Tools 4.9k | about 16-56k | 9/9 passes; fastest single run, more variance |

Direct MCP is richer; the CLI skill has basic set of MCP tools exposed and is useful when you want a shell-native path with progressive discovery and lower upfront tool-surface load.

pi is a good example of the target shape: a lean coding agent that can run commands but may not come with MCP mounted in the runtime. For that kind of agent, mcp2cli is less about replacing MCP and more about reaching MCP servers from the interface the agent already has.

Why

Native MCP clients expose tool schemas directly to the model. That is convenient, but it has an upfront context cost and assumes the agent runtime can mount MCP servers.

Some agents are shell-first. They can run commands but do not come with MCP installed. For those agents, mcp2cli is a practical bridge: the agent discovers browser tools through CLI commands (--list, --search, command help, snapshots, small probes) and keeps the working workflow in a skill.

This skill packages that workflow:

preflight local prerequisites before browser work
use one named browser session for repeated commands
prefer text snapshots and DOM assertions over screenshots
prefer form-level interactions over many one-off clicks/fills
inspect console and network output when the page misbehaves
stop the session when done

Install

From your project root (or add -g for a global install):

npx skills add maxim-saplin/chrome-devtools-mcp2cli

Manual alternative: copy everything under skills/chrome-devtools-mcp2cli/ into your agent's skills folder, for example:

.agents/skills/chrome-devtools-mcp2cli
.claude/skills/chrome-devtools-mcp2cli

Requirements

The target machine must have:

uvx
Node.js
npx
Google Chrome or Chrome for Testing available to chrome-devtools-mcp

Platform support

Tested here:

macOS

Expected, but not verified in this package:

Linux, assuming uvx, Node.js, npx, and Chrome/Chrome for Testing are installed and available on PATH

Not tested:

Windows native shells

The included preflight script is Bash. On Windows it may work through Git Bash or WSL if uvx, Node.js, npx, and Chrome launching all work from that environment, but this package does not claim Windows support yet.

The skill starts Chrome DevTools MCP with an isolated browser profile by default.

By default it uses the Chrome DevTools MCP README posture:

npx -y chrome-devtools-mcp@latest --isolated

Override MCP_SERVER_CMD if a run needs headless mode, pinned versions, disabled usage statistics, disabled CrUX lookups, custom Chrome paths, or connection to an existing debuggable browser.

Quick preflight

From the installed skill directory:

bash scripts/preflight.sh

Optional target smoke check:

bash scripts/preflight.sh http://localhost:8501 browser-preflight

Preflight checks browser-driving prerequisites. Target app reachability is not a mandatory preflight gate unless you pass a URL.

Repository layout

README.md
LICENSE
.gitattributes

skills/
  chrome-devtools-mcp2cli/
    SKILL.md
    scripts/
      preflight.sh

Use

Ask in plain language:

"Run a browser smoke test of the local app."
"Open the app, verify the main navigation, inspect console errors, and report evidence."
"Use a browser session to reproduce the failing interaction and capture page state."

The agent should run preflight first, keep one named session for repeated commands, and clean up the session at the end.

Security and privacy

Chrome DevTools MCP can inspect and modify browser state. Do not use this skill against pages containing secrets or personal data you do not want exposed to the agent transcript.

When passing credentials to any MCP or API server through mcp2cli, use environment variables or files rather than literal command-line secrets.

Chrome DevTools MCP collects usage statistics by default and performance tooling may use CrUX data unless disabled. Set MCP_SERVER_CMD with the upstream privacy flags when that matters for the run.

MCP Servers