Ask Claude Code why your build is slow and it will read your build config, because that is what it can see. It cannot see that the disk is at 96 percent, that a runaway process has been pinning four cores since yesterday, or that memory pressure has been red for an hour. The information that would answer the question is on the machine, and the agent has no way to reach it.
What MCP changes
The Model Context Protocol gives an agent a typed set of tools it can call. A server running locally can expose real system state as structured data rather than asking the model to parse shell output, which matters because parsing ps and df output is exactly where agents produce confident nonsense.
Burrow ships an MCP server with twenty-six tools. Fourteen are read-only and available immediately: snapshots of current system state, long-range history, top processes by CPU or memory, disk analysis, duplicate detection, listening ports, network throughput, and a health check. The tools that change the machine, cleaning and uninstalling among them, need their own explicit opt-in before an agent can call them at all.
Wiring it up
Point Claude Code at the app's MCP entry point:
{
"mcpServers": {
"burrow": {
"command": "/Applications/Burrow.app/Contents/MacOS/Burrow",
"args": ["--mcp"]
}
}
}There is also a loopback HTTP API on 127.0.0.1:9277 for tools that speak HTTP rather than MCP. Both are local only, and the HTTP surface can be turned off in Settings.
What it is actually good for
- Diagnosing a slowdown while it is happening, with the agent correlating CPU, memory pressure, disk, and network in one pass instead of asking you to run four commands.
- Answering questions about the past. History keeps ninety days of samples, so "what was hammering this machine at 2am" has an answer.
- Finding disk space before a large build or model download, with the agent able to see which directories are actually large.
- Pre-flighting an environment: checking whether a port is free, whether the disk has headroom, whether a previous run left processes behind.
The part that matters
The interesting property is not that an agent can run a cleanup. It is that the agent stops guessing about the machine. Most bad agent advice about performance comes from having no data and pattern-matching on the code instead, and a typed feed of real measurements removes the need to guess.
The full tool list, including which are gated, is in the agent documentation in the repository.