Impossible to start two Chrome instances from DevTools for agents MCP server when using Claude Code CLI

[gilbertococchi]

Description of the bug

It looks like it's not possible to start multiple sessions using DevTools for agents MCP server with Claude Code CLI.

Reproduction

• Open Claude Code CLI
• Insert prompt asking to test a page on localhost
• Claude Code CLI invocates DevTools for agents MCP server opening Chrome Browser instance.
• In the meantime, open another tab with a different prompt asking to browse another page
• CHrome is locked and the secont tab is not triggering a new Chrome instance.

Expectation

Claude Code CLI should be able to trigger multiple Chrome Browser instances, only just one per time

MCP configuration

No response

Chrome DevTools MCP version

0.26.0

Chrome version

No response

Coding agent version

No response

Model version

No response

Chat log

No response

Node version

No response

Operating system

None

Extra checklist

• I want to provide a PR to fix this bug

[natorion]

We discussed this when we worked on #1019. @OrKoN thoughts?

[OrKoN]

Yes, if ClaudeCode does not start a new instance and shares the instance between conversations, the --experimentalPageIdRouting is the way to make it work for concurrent conversations.

[natorion]

We should document this feature in our docs. Seems like we have nothing.

[OrKoN]

it is still experimental and we have not done evals. Let's bring to planning if it is important. Although with this bug I am not sure if the expectation is that we should manage multiple Chrome instances with one MCP server instance. We decided against this in #1019

[passtas]

@gilbertococchi, if you would start MCP with --isolated flag, you can run as many parallel individual sessions as you'd like even now. Each Claude session will use its own context when spinning up the mcp.

There are basically two features available atm:

• For "subagents" (who shares the tools from parent main agent) but within the same session - you can use pageId routing - so they will reuse the same browser context and parallel work will be managed in separate tabs
• For "Team of Agents"/"Swarm" or just two or more instances of the CLI - use the isolated flag - each session will use its own instance of the mcp and browser.

My default now is:

"chrome-devtools": {
            "command": "npx",
            "args": ["-y", "chrome-devtools-mcp@latest", "--experimentalPageIdRouting", "--isolated"]
        }

Basically it works everywhere enabling both modes of operation.