OpenLIT MCP server: read traces through ClickHouse

OpenLIT instruments MCP clients and servers so their tool calls show up as OpenTelemetry traces. To let an agent read those traces, connect it to the ClickHouse database OpenLIT writes to.

Checked 3 October 2026 · from OpenLIT's documentation index, SDK configuration and Docker Compose file, the openlit GitHub organization, and the official MCP registry

One of 91 guides to monitoring and developer tools. Each covers what an AI coding agent can get out of the tool through MCP, the protocol that lets the agent query the tool directly.

How OpenLIT works with MCP

OpenLIT is an open-source observability platform for LLM and agent applications, licensed Apache 2.0. Its SDK instruments model calls, vector databases and agent frameworks, then sends OpenTelemetry (OTel) data over OTLP, OTel's wire protocol.

MCP appears in OpenLIT as something to watch. Two pages titled "Get started with MCP Monitoring" instrument your MCP clients and servers so each tool call becomes a span. A third, Scanner connectors, scans GitHub repositories for MCP and agent-SDK findings.

A self-hosted install runs an OTLP receiver, a dashboard and ClickHouse, a columnar SQL database. The default Docker Compose file names the database openlit and exposes ClickHouse's HTTP port, 8123.

OpenLIT's REST API covers prompts, Vault secrets, the rule engine, model pricing, API keys and scanner jobs. It has no endpoint for querying traces.

Otter, OpenLIT's built-in chat assistant, does turn questions into read-only SQL over your telemetry. It runs inside the OpenLIT UI, so an external agent cannot call it.

Every vendor here gets the same four checks: official, maintained, distributed and usable. OpenLIT passes 0 of them.

All three rows resolve to one ClickHouse schema. An agent with SQL access can answer trace, cost and GPU questions, provided it knows the OpenTelemetry column names OpenLIT uses.

What can an AI agent do with OpenLIT?

Do you need the MCP server at all?

For trace and cost questions, mcp-clickhouse is the shortest route. It exposes run_query, list_databases and list_tables, and refuses writes unless CLICKHOUSE_ALLOW_WRITE_ACCESS is set to true.

If the SDK exports somewhere other than OpenLIT, query that backend instead. OpenLIT documents exporters for Grafana Cloud, Datadog, New Relic, Dynatrace, Elastic, SigNoz and Langfuse, among others.

For prompts, Vault secrets or model pricing, use OpenLIT's REST API. It returns prompts already compiled with your variables and pricing grouped by provider.

Use ClickHouse's MCP server for anything OpenLIT recorded, and OpenLIT's REST API for anything OpenLIT manages.

OpenLIT integration setup

OpenLIT itself needs no MCP configuration. The setups below point ClickHouse's MCP server at the database OpenLIT already writes to.

claude mcp add-json openlit-clickhouse '{"command":"uv",
  "args":["run","--with","mcp-clickhouse","--python","3.12","mcp-clickhouse"],
  "env":{"CLICKHOUSE_HOST":"localhost","CLICKHOUSE_PORT":"8123",
         "CLICKHOUSE_USER":"agent_ro","CLICKHOUSE_PASSWORD":"<password>",
         "CLICKHOUSE_SECURE":"false"}}'

{"mcpServers":{"openlit-clickhouse":{
  "command":"uv",
  "args":["run","--with","mcp-clickhouse","--python","3.12","mcp-clickhouse"],
  "env":{"CLICKHOUSE_HOST":"localhost","CLICKHOUSE_PORT":"8123",
    "CLICKHOUSE_USER":"agent_ro",
    "CLICKHOUSE_PASSWORD":"${env:OPENLIT_CH_PASSWORD}",
    "CLICKHOUSE_SECURE":"false"}}}}

Every other client

Each block below is the configuration for one client, with the file path and the key that client expects.

SELECT SpanAttributes['gen_ai.request.model'] AS model,
       round(sum(toFloat64OrZero(SpanAttributes['gen_ai.usage.cost'])), 4) AS usd
FROM openlit.otel_traces
WHERE Timestamp > now() - INTERVAL 7 DAY
GROUP BY model ORDER BY usd DESC LIMIT 20

Read-only access, permissions and security

• Create a dedicated ClickHouse user

  OpenLIT's Compose file defaults to the default user with password OPENLIT, which is far too broad for an agent.

• Leave writes off

  mcp-clickhouse keeps CLICKHOUSE_ALLOW_WRITE_ACCESS=false unless you change it.

• Expect sensitive text in spans

  Prompts and completions are captured by default, so grant the agent only the tables a task needs.

What the vendor documents

OpenLIT's quickstart says it "automatically instruments MCP alongside LLMs, VectorDBs, and frameworks by default". Its architecture diagram shows the SDK sending OTLP to the OpenLIT receiver, which stores data in ClickHouse for the UI to read.

Should you connect an agent to OpenLIT?

Yes, through ClickHouse, if your OpenLIT data stays in OpenLIT's own database. Best for self-hosted OpenLIT users who want an agent to answer trace and cost questions in SQL.

Strongest fit

Yes, through ClickHouse, if your OpenLIT data stays in OpenLIT's own database.

Main advantage

Traces, cost and GPU metrics share one ClickHouse schema that a SQL-capable agent can read.

Main weakness

The agent has to write SQL against OpenTelemetry tables, with no OpenLIT-specific tools to lean on.

Operational risk

Spans can hold full prompt and completion text, and a broad query hands it to the agent.

Frequently asked questions