RFC: CLI wrapper to enable agent-authored browser scripting

[natorion]

Is your feature request related to a problem? Please describe.

The current MCP interaction model requires an agent to perform a full read/think/act loop for every single browser action. For complex tasks—like scraping multiple pages or conditional navigation—this is inefficient, slow, and consumes excessive tokens.

Agents currently lack the ability to "batch" actions or apply simple logic (like loops or retries) without re-prompting the LLM for every step.

Describe the solution you'd like

I would like to explore and understand interest in chrome-devtools-cli: a command-line interface wrapper for the Chrome DevTools MCP server.

By exposing MCP capabilities via a CLI, agents can write and execute bash scripts to control the browser. This unlocks powerful workflows:

• Batching & latency reduction: Instead of 10 separate HTTP round-trips, an agent can generate a single script to perform a sequence of actions immediately.
• Logic & control flow: Agents can utilize standard bash logic (loops, conditionals) to handle tasks without constant re-prompting.

Example agent-authored script:

#!/bin/bash
urls=("https://example.com/page1" "https://example.com/page2")

for url in "${urls[@]}"; do
  chrome-devtools-cli navigate_page --url "$url"
  chrome-devtools-cli wait_for --text "Success" --timeout 2000
  chrome-devtools-cli take_screenshot --filePath "results/$(basename $url).png"
done

Describe alternatives you've considered

• Status quo (single tool calls): We could continue relying on individual tool calls, but this prevents optimization of latency and token usage for repetitive tasks.
• Generating Puppeteer/Playwright scripts: Agents could write raw Node.js scripts, but this requires a heavier runtime environment, and more complex code generation compared to a simple, pre-installed CLI wrapper.

Additional context

No response

[BogdanCerovac]

I like this suggestion - as mentioned in the alternatives - we don't have user and agent friendly alternatives at the moment (and CLI, when providing stable and robust interface, would perhaps also be useful as a scripting replacement outside of agent and LLM context).

To make things even more efficient, I wonder if we can also consider additional CLI feature that would run multiple instances in parallel?

I've been using https://github.com/thomasdondorf/puppeteer-cluster for such tasks and must say that it really speeds things up. I guess on Linux systems that is not an issue within bash, and I guess we can use ForEach-Object -Parallel on Windows PS (without WSL) but in both cases we need to do quite some additional checking etc. to make it work.

Perhaps it could already be abstracted / provided with a CLI feature?

[natorion]

I like this suggestion - as mentioned in the alternatives - we don't have user and agent friendly alternatives at the moment (and CLI, when providing stable and robust interface, would perhaps also be useful as a scripting replacement outside of agent and LLM context).

To make things even more efficient, I wonder if we can also consider additional CLI feature that would run multiple instances in parallel?

I've been using https://github.com/thomasdondorf/puppeteer-cluster for such tasks and must say that it really speeds things up. I guess on Linux systems that is not an issue within bash, and I guess we can use ForEach-Object -Parallel on Windows PS (without WSL) but in both cases we need to do quite some additional checking etc. to make it work.

Perhaps it could already be abstracted / provided with a CLI feature?

I think running multiple instances in paralell should already work with --isolated

[WebCloud]

This is a great option. Having an official CLI for devtools would be great.

Linking here for it is relevant:
https://github.com/vercel-labs/agent-browser