Observability for MCP servers

See every tool call your MCP server makes.

Lens reads the ClickHouse tables your server already writes and turns them into session timelines, latency, errors, and the assistant's own reflections — so you can make the server better.

Open dashboard Browse a session
Meta Claudesess_92ff86fd · 11 calls
select_ad_account602ms
get_write_action_schema401ms
execute_write_action1.00s
search_interest_candidates24.7s
preview_targeting_reach5.48s
execute_write_action26.3s

Called by the agents your users already run

ClaudeChatGPTGemini

Across any MCP provider

MetaGoogle

A trace you can actually read

Every tool call is an icon-led row sized by its real duration — the 24-second interest fetch dwarfs the 400ms schema lookup, at a glance. Click any call for inputs, outputs, and errors.

The assistant grades its own work

At session end the model writes what went wrong and what to fix, with priority. Lens turns those reflections into a cross-session board — the shortest path from a confusing tool to a server change.

Errors, grouped and traceable

Failures clustered by error code and tool, each linking back to the exact session that produced it. p50/p95 latency per tool, honest about what an MCP server can measure.

Open the lens.

Open dashboard