Use knotel

MCP and AI assistants

Connect Claude or another assistant and ask it about your traces.

knotel has a built-in Model Context Protocol server, so Claude and other AI assistants can explore your traces for you: "why did checkout get slower today?", "which Mongo queries are slowest?", "what's calling the payments service?".

Create an access token

In Settings → Access tokens & MCP, create a token and copy it; it's shown once. A token can read every project as you.

Connect

Claude Code

claude mcp add --transport http knotel https://YOUR-INSTANCE/mcp \
  --header "Authorization: Bearer knp_YOUR_TOKEN"

Other clients that support remote MCP servers over Streamable HTTP need the URL and header. For example, in a project's .mcp.json:

{
  "mcpServers": {
    "knotel": {
      "type": "http",
      "url": "https://YOUR-INSTANCE/mcp",
      "headers": { "Authorization": "Bearer knp_YOUR_TOKEN" }
    }
  }
}

Tools

Tool	What it answers
`list_projects`	Which projects exist.
`list_services`	Health per service: requests, errors, error rate, p50/p95/p99.
`list_fields`	Attribute keys on recent spans, most common first.
`run_query`	Filter, group by up to 3 fields, and aggregate. See Querying.
`timeseries`	Requests, errors and p95 over time.
`compare_periods`	The same query over two windows: what changed, per group. Grouped by service.version it reads a rollout; the baseline can be the window before, or the same hours a day or a week ago.
`find_traces`	Example traces matching filters, newest or slowest first.
`get_trace`	Every span of one trace as a tree, with attributes, events, links, self time and critical-path time, plus where the latency went, repeated queries, and the log lines written during the trace.
`search_logs`	Log records, newest first, by service, minimum severity, message substring, trace id or any log attribute.
`list_errors`	Error groups, most frequent first; with a group id, its recent occurrences and latest stack trace.
`find_outliers`	What the slowest (p95, p99 or a minimum duration) or failed spans have in common that the rest don't: fields ranked by how over-represented their top value is, with filters to drill in.
`list_agent_sessions`	AI agent conversations with their model calls, tool calls, tokens and cost; with a conversation id, one session step by step.
`list_sessions`	Browser sessions, or the people behind them with users=true.
`service_map`	Services, databases and hosts, and the calls between them, across all projects.

Tools take a project (id or name; optional when there's only one), and either a time_range (15m to 30d, default 1h, or 24h for the service map) or ISO 8601 start and end times.

Things to ask

"Give me a health check of every service in the last hour."
"p95 for the api service went up this afternoon. Find slow traces and tell me which span is responsible."
"Which MongoDB collections and operations are slowest this week?"
"Show error rate by HTTP status code for the checkout routes."
"Why are the slowest 1% of api requests slow today? Show me a trace for the top finding."
"What does the browser app call, and what do those services depend on?"

Tokens are powerful

An access token reads every project's spans, including URLs, queries and error messages. Give each client its own token so you can revoke one without affecting the others.