optimize: MCP finding suggests `claude mcp remove` for claude.ai connectors, which cannot remove them (and collides with same-named local servers)

[adamkwhite]

Problem

The "N MCP servers with low tool coverage" finding emits a remediation command per server:

-- Run this command ────────────────────────────────────
Remove underused servers, or trim their tools in your MCP config:
  claude mcp remove 'firecrawl-mcp'
  claude mcp remove 'claude_ai_Slack'
  claude mcp remove 'filesystem'
  claude mcp remove 'claude_ai_Google_Calendar'
  claude mcp remove 'claude_ai_Netlify'

claude mcp remove edits local MCP config (~/.claude.json / .mcp.json). The claude_ai_* entries are claude.ai connectors, provisioned server-side and toggled through /mcp or claude.ai's connector settings. They have no local mcpServers entry, so the command cannot remove them.

codeburn 0.9.20 (npm), Claude Code 2.1.228.

Evidence

Detection is correct — these servers really are loaded and unused. Only the remediation is wrong.

My local mcpServers at user scope, at the time codeburn suggested removing three claude_ai_* servers:

autoclaude-channel, claude-memory, google-workspace, playwright

No claude_ai_* key exists anywhere in ~/.claude.json or ~/.claude/settings.json. codeburn is deriving these names from transcript tool prefixes (mcp__claude_ai_Netlify__*), which is a reasonable detection method, but the fix assumes local config ownership that isn't there.

There is also a naming mismatch that would defeat a name-based match. Claude Code's own disabledMcpServers lists spell the connector with a dot and a space:

"disabledMcpServers": ["jira", "MongoDB", "claude.ai Jira",
  "claude.ai Google Calendar", "firecrawl-mcp",
  "claude.ai Slack", "claude.ai Netlify"]

claude.ai Netlify in config vs claude_ai_Netlify in the tool prefix.

Second problem: name collision with a real local server

I run a local netlify MCP, configured at project scope and actively used in two projects. codeburn's advice was claude mcp remove 'claude_ai_Netlify', referring to the unused connector.

Read quickly, that is "remove Netlify" — while a differently-scoped Netlify server is a live dependency. A user who acts on it either removes the wrong thing or spends time working out that these are two different servers with near-identical names.

Third, smaller: the finding is not scoped where it applies

Connectors are enabled/disabled per project. In my case the connector was already disabled in one project and still loading in another, so the correct action was a per-project toggle, not a global remove. The finding aggregates across projects and suggests one global command.

(Related: #974 covers a separate inflation in the session population these findings run over.)

Suggested direction

Listing options rather than assuming a preference:

1. Detect the claude_ai_* prefix and emit connector-appropriate guidance — "disable in /mcp, or at claude.ai → Settings → Connectors" — instead of claude mcp remove.
2. Before suggesting claude mcp remove <name>, confirm <name> actually resolves in local MCP config; if it doesn't, say where the server is actually configured rather than emitting a command that will fail.
3. When an unused server's name is a substring/variant of a used server elsewhere, disambiguate explicitly in the finding text (scope + config source), so the advice can't be misread as targeting the live one.
4. Where per-project usage data already exists, scope the suggestion to the projects where the server is loaded-but-unused.

Option 2 generalises best — it turns any "command targets config codeburn doesn't own" case into a diagnostic rather than a broken command.