Browser evidence is the missing acceptance layer for AI coding agents

Today's Doramagic metrics had a clear signal: `doramagic-chrome-devtools-mcp-pack` led the set with 120 clones, 60 unique cloners, and 44 views.

That is interesting because it is not a broad agent framework. It is a browser-verification capability pack.

AI coding agents often look strong while reading and editing code, then become weaker at the browser boundary. The failure is usually not dramatic. It looks like this:

- a page loads blank, but the agent reports completion;
- a button exists, but the real click flow fails;
- console errors are ignored;
- a network request fails, but the agent only reasons from source files;
- screenshots are collected without being tied to acceptance criteria.

Chrome DevTools MCP-style access helps because it turns browser state into evidence the agent can inspect.

The Doramagic angle is to package that evidence loop into something a user can carry into a host like Codex, Claude Code, Cursor, or Aider:

```text
open the URL
inspect console
run the key interaction
record visible state
compare against acceptance criteria
repair or stop
```

The important part is the stop condition. A browser tool should not make the agent more confident by default. It should make the agent more accountable.

Pack:

https://github.com/tangweigang-jpg/doramagic-chrome-devtools-mcp-pack

This is an unofficial AI capability pack prepared by Doramagic. Unless the upstream project explicitly says otherwise, it is not an official upstream release.