Browser extension

[pdchemist15]

The browser extension in either Chrome or Edge- Cannot see tools. Also the server cannot see the AI Chat. Ive installed reinstalled and the instructions say that after setting up and opening an AI window that AI will finish setting up. I can only see network tools not local tools. Calude desktop and Fusion 360 add-i is working well. I feel like I'm missing and not inderstanding something in the setup. Claude when using browsers can see the servers and connect there are just no tools available. Let me know if there is anything i can add to help troubleshoot.

[pdchemist15]

ive narrowed this down i think to a Claude extension incompatibility- Claude Desktop works fine- Chat GPT works fine through the Browser extension. Any suggestions? Claude cannot work through the browser. its interesting how ChatGPT has no issues. any thing i can provide to help resolve this.

[agrrobert]

First let me state that the local mcp server connected to Claude Desktop works beautifully.

The trouble I am having is with the browser extension, and perhaps I missed a step. It installed correctly in Chrome on Windows 11 and I added my authorization key. In order to facilitate debugging I checked the box to auto-approve everything.

The symptoms:

Claude Web doesn’t work at all. Nothing gets injected and it doesn’t know that mcp-link exists.

        **use the browser's mcp-link tool to open a new tab and go to example.com**
        2:22 PM
        Assessed unavailable tool and prepared honest clarification
        I don't have access to a browser MCP tool or the ability to open browser tabs. 
       That's not one of the tools available to me in this environment.

ChatGPT 5.2 is hit and miss: When I open a new chat the prompt gets injected and it was able to list my tabs. (though it took over 3 minutes for it think through the process.)
The AI seems to struggle with a lack of clarity in the tool instructions. I another chat I asked it to count my tabs and it was never able to complete. The thinking log is quite long, but I’ll include some snippets:

mcp-link system prompt.md
File
can you use the browser mcp-link tool to count how many tabs i have open?

Reading documents
Reading documents
Explaining MCP tool limitations
We can explain that while I don't have direct access to the MCP browser tool in this chat, it’s available on the user’s local setup.
I can guide them to run the operation and count the returned list or they can upload the output here for further analysis.

Crafting the payload and resolving conflicts
I’m noticing a discrepancy in the documentation—operation is being reused inside the payload, which causes a conflict because
JSON can’t have duplicate keys. I’ll need to resolve this.

Inspecting and resolving duplication in the file
I think the key might be "operation" or "browser_operation," but it seems like there’s some corruption with duplicated fields.
I’ll need to carefully inspect the file content through a search, checking for specific terms like "tool_unlock_token" and "LIST_TABS."
I'll also need line numbers to pinpoint where the corruption happens. Let’s dig into the file.

Investigating file search results

  - There’s a bit of a mystery here. The search didn’t return results even though it showed a file citation earlier. It could be that
  - the file wasn't indexed properly or that the contents didn’t match the query. I might try to access the file directly using the given path to get the information needed.
  - There’s some confusion around the citation format. The system instruction requires a line number, 
  - but I know I can also include the file citation as a link
  - The tool call should return a list of tabs, but since our environment doesn't allow executing the tool, we can't both call it and count in the same turn.
  - The file's schema on "tool reply.txt" leaves out key details about the exact keys needed for the "execute" call, especially regarding the operation field
  -  I think I can attempt an "execute" operation with specific parameters, but the ambiguity of these tools throws a challenge
  - Trying to figure out the correct tool call structure
  - There's a bit of back-and-forth here about avoiding duplication in the JSON object. 
  - I think they might’ve accidentally used "operation" instead of the intended "browserOperation" for the operation key. 

It continues struggling like this for quite a while, then finally generates the XML and then stalls.

<zntml:invoke name="mcp_ragtag_aura_browser">
<zntml:parameter name="input">
{ "operation": "execute", "tool_unlock_token": "e5076d", "browser_operation": "LIST_TABS" }
</zntml:parameter>
</zntml:invoke>
</zntml:function_calls>

Nothing is returned. Similar results when asking it to execute fusion commands—a lot of thrashing and confusion and then no result.

Interestingly, ChatGPT 5.1 Instant had no trouble and counted the tabs quickly. It still was not able to complete a simple Fusion task.
Perhaps the newer model “overthinks” and needs more clarification. Just like many smart people..