MCP observability tool adds per-tool health grades to catch silent agent failures
Thirumalaiboobathi · reddit · 2026-10-07
The problem: MCP tools can return isError: true inside a successful response, so standard OpenTelemetry shows a green span while the agent is actually failing. The new opentel-mcp-ui release tackles this head-on:
- Per-tool health grades (A–F) based on silent failures, errors, retry loops and latency, with explanatory tooltips
- 'Errors your agent can't act on' flag: catches failures returning empty or near-empty error text the agent can't recover from (checks length only, never content)
- Optional coverage for resources and prompts, not just tools
- Metrics flush on exit (stdin close, SIGTERM, SIGINT)
- Fixed SDK v2 McpServer instrumentation when declaring the tools capability
Try it with npx opentel-mcp-ui --demo. Open source on GitHub, with opentel-mcp and opentel-mcp-ui npm packages.
More from coding & agent
- Agent Evals 101: Benchmarking and Regression-Testing Agent Harnesses, 35B vs 120B — blaizedsouza · 2026-10-07
- Open-source docbrain: a self-hosted context engine that flags outage risks in your IDE — abhipsnl · 2026-10-07
- A weekly security audit workflow for AI agents: supply chain, prompt injection, MCP and more — alexcovo_eth · 2026-10-07
- Besxar builds Fabcast console to auto-process flight test data in-house — AviatorAshley · 2026-10-07
- Graphlit shuts down after 5 years: working product, failed business — KirkNewcombe · 2026-10-07
- One MCP server, mixed results: HTML UI support varies wildly across hosts — dfinke · 2026-10-07