Aggregate calls to the hosted MCP server at
/mcp. No per-user identities — counts only. Read tools need no
sign-in, so most calls have no identity behind them at all.
Each step of the OAuth sign-in an MCP client runs through GitHub, in order. Where the numbers drop between two steps is where people or agents give up before calling a single tool. Counts only, never who.
Every call that did not succeed, over MCP and REST, by where it
happened and its error code — so it shows where agents get stuck, not only where
they get through. Most are the service working as designed (a malformed hash, a
daily limit). internal_error is a bug on our side.
Agents can tell us what went wrong through send_feedback.
Only the count is shown here; a person reads the messages.