MCP/Tools Health

The MCP/Tools tab provides health metrics for every MCP server and built-in tool used across your team's sessions.

Tool Roster

Two views:

MCP Servers — every MCP server seen across your team's sessions, with per-tool drill-in
Built-in Tools — native agent tools (file read/write, shell, search) and shell binaries promoted to individual rows

Tuneloop MCP/Tools tab — error rate over time chart, tool roster with call counts, sessions, and error rates

Per-Tool Metrics

For each tool or server:

Call count and session count
Error rate and empty-result rate
Error rate trend over time
Per-repo usage breakdown
Session call list with transcript deep-links

Tuneloop MCP server detail view — Atlassian server showing tool breakdown, call count, sessions, error rate, and empty results

Error Categories

Every tool error is automatically classified into one of 19 categories using deterministic pattern matching — no LLM required. Categories include:

Precondition — tool preconditions not met (e.g., edit before read, stale file)
Not found — file, path, or resource missing
Permission — OS or filesystem access denied
Policy blocked — harness sandbox or approval policy blocked the action
User declined — user declined an approval prompt
Test failure — test run reported failures
Compile error — syntax errors, missing modules, type errors
Network — connection refused, DNS failure
Rate limit — 429 or quota exceeded
Integration error — MCP server transport-level failure

Classification is harness-agnostic — Claude Code, Codex, and OpenCode error strings all map into the same buckets. Tools with high error rates surface a dominant error category to help you diagnose the root cause quickly.

Error Attribution

For shell commands, Tuneloop attributes failures to the specific binary in a compound command — so npm install | grep error blames npm, not grep.