MCP server: debug in your editor

Vorlo ships an MCP server, so your AI coding assistant can read your agent's diagnoses directly. Ask "why did my last run fail?" in Claude Code or Cursor and get the root cause, the fix, and the failing step — without opening a dashboard.

Setup

The MCP server ships inside the SDK on both registries — use whichever runtime you already have.

Claude Code (Node)

claude mcp add vorlo --env VORLO_API_KEY=vrlo_... -- npx -y -p vorlo-trace vorlo-mcp

Claude Code (Python)

pip install vorlo-trace
claude mcp add vorlo --env VORLO_API_KEY=vrlo_... -- vorlo-mcp
# or without installing first:
claude mcp add vorlo --env VORLO_API_KEY=vrlo_... -- uvx --from vorlo-trace vorlo-mcp

Cursor / other MCP clients

Add to your MCP configuration:

{
  "mcpServers": {
    "vorlo": {
      "command": "npx",
      "args": ["-y", "-p", "vorlo-trace", "vorlo-mcp"],
      "env": { "VORLO_API_KEY": "vrlo_..." }
    }
  }
}

Tools

Tool	What it does
`why_did_my_last_run_fail`	The headline: fetches your most recent failed run and returns the diagnosis — root cause, fix, confidence, and every step with status and latency.
`get_session_diagnosis`	Full diagnosis + step summary for a specific session id.
`list_recent_sessions`	Recent runs, optionally filtered by status (failed / success / running).
`get_failure_clusters`	Failures grouped by root cause over a lookback window — fix the pattern, not the symptom.

The workflow this unlocks

Your agent fails in production. A Slack alert fires with the diagnosis.
You open your editor and ask: "why did my last run fail, and fix it".
Your assistant pulls the Vorlo diagnosis (root cause + fix hint), finds the code, and applies the fix — the whole loop without leaving the editor.

The MCP server is read-only and uses your API key from the environment — it can read diagnoses, never modify anything. Requires vorlo-trace 0.3.0+ (npm or PyPI).

Quickstart: JS / TS →Instrument your agent first — two lines of code.API & configuration →Init options, env vars, ingest contract.