# MCP Server. Telemetry Your Agents Can Query. > Connect AI agents directly to your telemetry with the OpenObserve MCP Server. Enable Claude and GPT-4 to autonomously query logs, metrics, and traces. Source: https://openobserve.ai/mcp-server/ --- Give your AI agents direct, programmable access to petabyte-scale telemetry. - [Start Free Cloud Trial](https://cloud.openobserve.ai/web/login/) - [Talk to a Human](/demo/) ### Autonomous RCA Pre-built tools let AI autonomously investigate logs, metrics, and traces instantly. ### LLM-Agnostic (BYOAI) Standardized via MCP. Use any LLM while your telemetry data stays secure. ### Zero-Trust AI Access Every agent query is authenticated, RBAC-scoped, and logged with complete audit trails. ## Give Your Agents the Same Access You Have ### Intent-Driven Telemetry Interrogation - **Natural Language to Optimized SQL/PromQL** - The MCP server translates intent into highly optimized SQL and PromQL queries, executing them instantly across your data streams. - **Autonomous Investigation** - Ask a question and it pulls latency metrics, distributed traces, and pod error logs to synthesize an RCA summary. ### Complete Observability Stack as Code - **Dynamic Tool Search (Context-Saving)** - Agents use tool_search to find relevant tools by keyword and fetch schemas just-in-time, executing via tool_call, maximum telemetry, zero token waste. - **Context-Aware Correlation** - Since logs, metrics, and traces are unified, the MCP server provides perfectly correlated context. ### Seamless Gateway & Client Integration - **Enterprise Gateway Ready** - Plug directly into enterprise MCP gateways like Cloudflare, Kong, or ContextForge. - **Direct Client Connection** - Connect Claude CLI or your IDE directly to the OpenObserve MCP server via stdio/SSE, executing telemetry queries locally in under 5 minutes. ### Enterprise-Grade Governance - **Two-Tiered Policy Enforcement** - Your MCP gateway defines which agents can access the server, while OpenObserve’s native RBAC restricts exactly which streams and queries those identities can execute. - **Immutable Tool Auditing** - Every MCP interaction logs the timestamp, agent ID, and accessed payload, providing SOC2-compliant audit trails. ## Measured against industry leaders Same telemetry, same workloads, one platform. Every number is OpenObserve against a named vendor - not an industry average. 8x cost reduction means you can unify your observability into a single platform. 140x storage means longer retention doesn't necessarily mean expensive bills. 10x performance means dashboards and queries load in milliseconds, not minutes. 30x compute efficiency means you index everything and don't have to compromise on sampling. See how much you would save switching today. - [See all comparisons](/comparison/) ## Teams trust OpenObserve to power their agents ## MCP Server FAQs ### How does the OpenObserve MCP server integrate with my infrastructure? Register the OpenObserve endpoint in your MCP client or gateway. When an LLM uses tool calling to request telemetry, the gateway securely routes the request to OpenObserve, which executes the query and returns structured data directly back to the model's context window. Direct connections work over stdio or SSE - Claude CLI or an IDE can be querying telemetry locally in under five minutes - while enterprise deployments can route through MCP gateways such as Cloudflare, Kong, or ContextForge. Governance is two-tiered: the gateway decides which agents may reach the server at all, and OpenObserve's native RBAC restricts exactly which streams and queries each identity can execute, with every interaction logged - timestamp, agent ID, and accessed payload - for SOC 2-compliant auditing. ### Can I enforce data residency and use local LLMs? Yes. The MCP protocol decouples data from inference: your telemetry stays in OpenObserve while the reasoning model runs wherever you choose. Point the server at self-hosted LLMs (like Llama 3) or enterprise-managed instances inside your VPC, and prompts, queries, and results never leave your network - maintaining strict data residency compliance. Because OpenObserve is LLM-agnostic, the same MCP tools work with Claude, GPT-4, or a local model; switching providers is a configuration change, not a re-integration. Combined with RBAC scoping and immutable audit logs on every agent query, this gives regulated teams autonomous AI investigation without surrendering control of where observability data lives or which model is allowed to see it. ### What is the latency overhead of using natural language for querying? Inference adds ~200-500ms for natural language translation. Once translated, queries execute at native, highly-optimized speeds. Standard SQL remains available for rapid troubleshooting, while LLMs save minutes of manual query writing for complex RCA. ### Do we need to instrument our apps differently for the MCP Server? No. OpenObserve supports standard instrumentation frameworks like OpenTelemetry (OTel). The MCP server is purely a read-path interface, leaving your existing collectors and telemetry pipelines completely untouched. ### How do we prevent AI hallucinations from crashing our observability backend? OpenObserve enforces strict rate limiting, query timeouts, and read-only tool execution. Destructive actions are not exposed, ensuring erratic agent behavior cannot impact system stability ## Explore guides, videos, and articles to help you get the most out of MCP Server. ### Model Context Protocol (MCP) Overview [Learn more](https://openobserve.ai/docs/integration/ai/mcp/) ### Integrate MCP with Claude CLI [Learn more](https://openobserve.ai/webinars-videos/integration-with-ai-tools-a-step-by-step-guide-using-mcp/) ### MCP Servers for Observability [Learn more](https://openobserve.ai/blog/mcp-servers-observability-guide/) - [Explore All Blogs](/blog/) ## Ready to get started? Try OpenObserve today for more efficient and performant observability. - Get Started For Free - [Schedule Demo](/demo/)