MCP Audit

The Problem

MCP tools and servers often generate hidden token overhead—from schema size, payload spikes, and inefficient tool patterns. These issues cause:

• Early auto-compaction — sessions end prematurely
• Slow agent performance — large contexts increase latency
• Unexpected cost increases — tokens add up faster than expected
• Context window exhaustion — hitting limits before finishing work

Who It’s For

🛠️ The Builder	💻 The Vibecoder
”Is my MCP server too heavy?"	"Why did my session auto-compact?”
You build MCP servers and want visibility into token consumption.	You use Claude Code daily and hit context limits without knowing why.
You need: Per-tool token breakdown, usage trends.	You need: Real-time cost tracking, session telemetry.

Demo

Real-time token tracking & MCP tool profiling — understand exactly where your tokens go.

What to Look For

Once you’re running MCP Audit, watch for these patterns:

1. Context Tax — Session starts with 10k+ tokens before you type anything (large schemas)
2. Payload Spike — Single tool call consumes far more tokens than expected
3. Zombie Tool — Tool in schema but never called (wasting schema tokens)
4. Auto-Compaction Trigger — Conversation compacts unexpectedly early

Transparency

What MCP Audit accesses: Your local session logs and artifacts only.

What Little Bear Apps sees: Nothing. All data stays on your machine.

What third parties see: PyPI sees download stats. GitHub sees repo activity. That’s it.

Network access: By default, fetches model pricing from LiteLLM API (cached 24h). No usage data is sent. Can be disabled in config.