Query your telemetry with MCP

Boidra Platform ships an MCP server that turns your agent-observability data into tools an AI assistant can call. Instead of clicking through the dashboard, you ask an assistant (Claude Desktop, Cursor, any MCP client) questions like "which agents errored in the last hour and what did they call?" and it queries your telemetry directly.

It reads the dashboard over the HTTP API (read-only by default) and never touches the database, so it works against any instance — local or a deployed SAP BTP dashboard.

What it can answer

Tool	Answers
get_overview	Fleet health over a window — requests, errors, latency, tokens, top services
list_services	Every agent + MCP tool with counts, error rate, tokens
search_traces	Traces by service / status / tool / min tokens-latency / window
get_trace	One trace's full span tree, tool input/output, and the agent's response
list_errors	Recent failures with the offending service and message
cost_breakdown · token_usage	Spend and tokens by model or service
top_prompts	Most frequent prompt terms
model_catalog	AI Core models nearing deprecation / retirement
list_alerts	Configured rules and which are firing now
list_judges · list_datasets · get_eval_run	Evaluation config and run status
evaluate_trace · run_dataset_evaluation	Run judges (opt-in; hidden in read-only mode)

Every tool is read-only except the last two. Set MCP_READONLY=1 to hide them entirely.

Connect Claude Desktop (local)

Build the server, then add it to your Claude Desktop config (%APPDATA%\Claude\claude_desktop_config.json on Windows, ~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "boidtelemetry": {
      "command": "node",
      "args": ["<path>/mcp-observability/dist/index.js"],
      "env": { "BOIDTELEMETRY_URL": "http://localhost:3010", "MCP_READONLY": "1" }
    }
  }
}

Restart Claude Desktop and ask: "Give me a Boidra Platform overview for this month", or "Which model is closest to retirement?".

Hosted on SAP BTP

The server also runs as a streamable-HTTP endpoint (POST /mcp), stateless so it scales horizontally on Cloud Foundry. Deploy with the bundled manifest.yml + xs-security.json — XSUAA gates the route with Reader and Evaluator scopes. Off-BTP, a static MCP_BEARER_TOKEN provides a simpler gate. This mirrors the hosted + local model of platforms like Bronto: run it yourself locally, or host it for the team.

See the server's README.md for the full configuration.

Ask AI — chat inside the dashboard

The dashboard also has a built-in Ask AI page (in the sidebar) that uses this same MCP server. You type a question in natural language; an LLM (SAP AI Core) answers by calling the observability tools, and the page streams the reply with every tool call shown as an expandable chip — so the chat is itself observable.

The loop runs on the dashboard backend (POST /api/chat), so LLM credentials stay server-side. It reaches the MCP server in HTTP mode (MCP_URL, default http://localhost:7010/mcp) and uses the same SAP AI Core configuration as Judges and Reporting. Set CHAT_LLM=mock to try the page without AI Core credentials — a deterministic keyword-routed model still exercises the real MCP tools.