--- title: LLM Observability MCP and Skills description: >- Connect AI agents to your LLM Observability traces and experiments using the Datadog MCP Server. breadcrumbs: Docs > LLM Observability > LLM Observability MCP and Skills --- # LLM Observability MCP and Skills {% callout %} # Important note for users on the following Datadog sites: app.ddog-gov.com, us2.ddog-gov.com {% alert level="danger" %} This product is not supported for your selected [Datadog site](https://docs.datadoghq.com/getting_started/site.md). (). {% /alert %} {% /callout %} ## Overview{% #overview %} The [Datadog MCP Server](https://docs.datadoghq.com/bits_ai/mcp_server/setup.md) enables AI agents to access your [LLM Observability](https://docs.datadoghq.com/llm_observability.md) data through the Model Context Protocol (MCP). The `llmobs` toolset provides tools for searching and analyzing traces, inspecting span details and content, and evaluating experiment results directly from AI-powered clients like Cursor, Claude Code, or OpenAI Codex. ## Setup{% #setup %} Connect an MCP-compatible client to the Datadog MCP Server with the `llmobs` toolset enabled. {% alert level="info" %} For full setup instructions, including Cursor and VS Code extension configuration, see [Set up the Datadog MCP Server](https://docs.datadoghq.com/bits_ai/mcp_server/setup.md). {% /alert %} ### Prerequisites{% #prerequisites %} - A Datadog account with permission to access LLM Observability data. - An MCP-compatible client (for example, Claude Code, Codex CLI, Cursor, Gemini CLI, or Kiro CLI). ### Endpoint{% #endpoint %} The MCP Server endpoint depends on your [Datadog site](https://docs.datadoghq.com/getting_started/site.md). Use the Datadog Site selector to display the endpoint for your site. Append `?toolsets=llmobs,core` to enable the LLM Observability and core toolsets. {% callout %} # Important note for users on the following Datadog sites: app.datadoghq.com, us3.datadoghq.com, us5.datadoghq.com, app.datadoghq.eu, ap1.datadoghq.com, ap2.datadoghq.com Endpoint for your selected site (): ``` ?toolsets=llmobs,core ``` {% /callout %} {% callout %} # Important note for users on the following Datadog sites: app.ddog-gov.com, us2.ddog-gov.com {% alert level="danger" %} This product is not supported for your selected site (). {% /alert %} {% /callout %} ### Connect{% #connect %} Choose remote authentication when possible. Use local binary authentication if your environment blocks the remote OAuth flow. {% tab title="Remote authentication" %} {% callout %} # Important note for users on the following Datadog sites: app.datadoghq.com, us3.datadoghq.com, us5.datadoghq.com, app.datadoghq.eu, ap1.datadoghq.com, ap2.datadoghq.com Remote authentication uses the MCP specification's [Streamable HTTP](https://modelcontextprotocol.io/specification/2025-03-26/basic/transports#streamable-http) transport. **Claude Code** (command line): ``` claude mcp add --transport http datadog-mcp "?toolsets=llmobs,core" ``` **Codex CLI** (`~/.codex/config.toml`): ``` [mcp_servers.datadog] url = "?toolsets=llmobs,core" ``` After adding the configuration, run `codex mcp login datadog` to complete the OAuth flow. **Gemini CLI, Kiro CLI, and other MCP-compatible clients**: ``` { "mcpServers": { "datadog": { "type": "http", "url": "?toolsets=llmobs,core" } } } ``` {% /callout %} {% callout %} # Important note for users on the following Datadog sites: app.ddog-gov.com, us2.ddog-gov.com {% alert level="danger" %} This product is not supported for your selected site (). {% /alert %} {% /callout %} {% /tab %} {% tab title="Local binary authentication" %} Local binary authentication uses the MCP specification's [stdio](https://modelcontextprotocol.io/specification/2025-03-26/basic/transports#stdio) transport. Use this method if remote authentication is unavailable. 1. Install the Datadog MCP Server binary: ```bash curl -sSL https://coterm.datadoghq.com/mcp-cli/install.sh | bash ``` The binary installs to `~/.local/bin/datadog_mcp_cli`. 1. Complete the OAuth login flow: ```bash datadog_mcp_cli login ``` 1. Configure your AI client. For Claude Code, add the following to `~/.claude.json`, replacing `` in the command path: ```json { "mcpServers": { "datadog": { "type": "stdio", "command": "/Users//.local/bin/datadog_mcp_cli", "args": [], "env": {} } } } ``` Alternatively, add the server with the Claude Code CLI: ```bash claude mcp add datadog --scope user -- ~/.local/bin/datadog_mcp_cli ``` {% /tab %} ### Authenticate with API keys{% #authenticate-with-api-keys %} The MCP Server uses OAuth 2.0 by default. If OAuth is unavailable, send a Datadog [API key and application key](https://docs.datadoghq.com/account_management/api-app-keys.md) as the `DD_API_KEY` and `DD_APPLICATION_KEY` HTTP headers: {% callout %} # Important note for users on the following Datadog sites: app.datadoghq.com, us3.datadoghq.com, us5.datadoghq.com, app.datadoghq.eu, ap1.datadoghq.com, ap2.datadoghq.com ``` { "mcpServers": { "datadog": { "type": "http", "url": "?toolsets=llmobs,core", "headers": { "DD_API_KEY": "", "DD_APPLICATION_KEY": "" } } } } ``` {% /callout %} {% callout %} # Important note for users on the following Datadog sites: app.ddog-gov.com, us2.ddog-gov.com {% alert level="danger" %} This product is not supported for your selected site (). {% /alert %} {% /callout %} For security, scope the API key and application key to a [service account](https://docs.datadoghq.com/account_management/org_settings/service_accounts.md) with only the required permissions. ## Agent skills{% #agent-skills %} Agent skills are prebuilt instruction sets for AI coding agents that automate common LLM Observability workflows. The `dd-llmo` skill set is available in the [Datadog agent-skills](https://github.com/datadog-labs/agent-skills) repository. It provides six skills for classifying sessions, diagnosing failures, analyzing experiments, generating experiment code with the `ddtrace.llmobs` SDK, and bootstrapping evaluators against your live production data. ### Install{% #install %} Install the `dd-llmo` skills with the following command: ```shell npx skills add datadog-labs/agent-skills --skill dd-llmo --full-depth -y ``` The skills require the `llmobs` MCP toolset to be connected. If you have not already connected it, run: ```shell claude mcp add --scope user --transport http "datadog-llmo-mcp" \ 'https://mcp.datadoghq.com/api/unstable/mcp-server/mcp?toolsets=llmobs' ``` Restart Claude Code after running both commands for the skills to appear. ### Available skills{% #available-skills %} | Skill | Invoke with | What it does | | ------------------------- | ---------------------------------- | -------------------------------------------------------------------------- | | Session classify | `/llm-obs-session-classify` | Classifies whether user intent was satisfied in a session, trace, or batch | | Trace RCA | `/llm-obs-trace-rca` | Root cause analysis on failing production traces | | Experiment analyzer | `/llm-obs-experiment-analyzer` | Analyze and compare LLM experiment results | | Experiment Python codegen | `/llm-obs-experiment-py-bootstrap` | Generate Python experiment code using the `ddtrace.llmobs` SDK | | Eval bootstrap | `/llm-obs-eval-bootstrap` | Generate evaluator code or publish online LLM-judge evaluators | | Eval pipeline | `/llm-obs-eval-pipeline` | End-to-end pipeline: classify → RCA → bootstrap evaluators | #### Session classification{% #session-classification %} `/llm-obs-session-classify` classifies whether user intent was satisfied in a given interaction. It draws from up to three signal sources: LLM Observability traces, RUM behavioral data, and Audit Trail events. The skill returns a `yes / partial / no` verdict with supporting evidence. Confidence improves with each additional signal source. ``` /llm-obs-session-classify session_id= /llm-obs-session-classify trace_id= /llm-obs-session-classify ml_app=my-chatbot --timeframe now-7d ``` #### Trace root cause analysis{% #trace-root-cause-analysis %} `/llm-obs-trace-rca` diagnoses why an LLM application is producing poor results. It selects an analysis mode based on the strongest available signal (LLM-judge eval verdicts, runtime errors, or structural anomalies) and compiles a structured RCA report. The report includes a failure taxonomy and concrete `BEFORE` / `AFTER` fix proposals grounded in trace evidence. When Claude Code has access to your codebase, the skill can search for the relevant source files and propose diffs inline. ``` /llm-obs-trace-rca ml_app=my-chatbot /llm-obs-trace-rca ml_app=my-chatbot eval_name=faithfulness --timeframe now-24h ``` #### Evaluator bootstrap{% #evaluator-bootstrap %} `/llm-obs-eval-bootstrap` analyzes production traces and proposes a suite of evaluators targeting the observed failure modes. It outputs one of three artifacts: Python `BaseEvaluator` / `LLMJudge` classes for offline experiments, a framework-agnostic JSON spec, or online LLM-judge evaluators published directly to Datadog. ``` /llm-obs-eval-bootstrap ml_app=my-chatbot /llm-obs-eval-bootstrap ml_app=my-chatbot --publish /llm-obs-eval-bootstrap ml_app=my-chatbot --data-only ``` #### Experiment analyzer{% #experiment-analyzer %} `/llm-obs-experiment-analyzer` retrieves experiment results and surfaces what changed between a candidate and a baseline: which metrics improved, which regressed, and where the candidate underperformed. ``` /llm-obs-experiment-analyzer experiment_id= /llm-obs-experiment-analyzer experiment_id= baseline_id= ``` #### Generate experiment code with the Python SDK{% #generate-experiment-code-with-the-python-sdk %} `/llm-obs-experiment-py-bootstrap` generates a self-contained Python experiment client that uses the `ddtrace.llmobs` SDK. The output is either a runnable `.py` script or a Jupyter `.ipynb` notebook matching the canonical reference notebook style. The dataset can come from a local JSON or CSV file, an existing Datadog dataset fetched by name, or a built-in inline sample. Every generated experiment is tagged with `generated_by=claude-code` so you can identify and filter Claude-generated experiments in the LLM Experiments list. ``` /llm-obs-experiment-py-bootstrap /llm-obs-experiment-py-bootstrap --dataset ./data/qa.json --format ipynb /llm-obs-experiment-py-bootstrap --dataset-name --project-name ``` #### End-to-end eval pipeline{% #end-to-end-eval-pipeline %} `/llm-obs-eval-pipeline` chains session classification, trace RCA, and evaluator bootstrap into a single supervised workflow with user checkpoints between phases. It is the recommended starting point when you have no existing evaluators for an application. ``` /llm-obs-eval-pipeline my-chatbot /llm-obs-eval-pipeline my-chatbot --timeframe now-30d --publish ``` For a complete guide to these skills and a recommended end-to-end workflow, see [Analyze LLM Applications with Claude Code Skills](https://docs.datadoghq.com/llm_observability/guide/claude_code_skills.md). ## Use cases{% #use-cases %} The LLM Observability MCP tools enable AI-assisted workflows for: - **Debugging agent execution**: Search for traces by ML app, error status, or custom tags, then examine span hierarchies and content to identify failures. - **Analyzing trace structure**: Visualize the full span tree of a trace to understand how agents, LLMs, tools, and retrievals interact. - **Investigating agent loops**: Review an agent's step-by-step execution loop to understand decision-making and tool invocation patterns. - **Evaluating experiments**: Get summary statistics for experiment metrics, compare results across dimension segments, and inspect individual events. - **Discovering experiment patterns**: Filter and sort experiment events by metric performance to find the best and worst-performing cases. - **Managing evaluators**: List, inspect, create, update, and delete evaluator configurations across an ML application or the entire organization. ## Available tools{% #available-tools %} The `llmobs` toolset includes the following tools: ### Trace and span tools{% #trace-and-span-tools %} {% dl %} {% dt %} `search_llmobs_spans` {% /dt %} {% dd %} Search for spans matching filters or a raw query. {% /dd %} {% dt %} `get_llmobs_trace` {% /dt %} {% dd %} Get the full structure of a trace as a span hierarchy tree, including span counts by kind, error indicators, and total duration. {% /dd %} {% dt %} `get_llmobs_span_details` {% /dt %} {% dd %} Get detailed metadata for one or more spans, including timing, error info, LLM details (model, token counts), metrics, and evaluations. {% /dd %} {% dt %} `get_llmobs_span_content` {% /dt %} {% dd %} Retrieve the actual content of a span field (input, output, messages, documents, or metadata) with optional JSONPath extraction. {% /dd %} {% dt %} `find_llmobs_error_spans` {% /dt %} {% dd %} Find all error spans in a trace with propagation context, grouped by span kind with error messages and stack traces. {% /dd %} {% dt %} `expand_llmobs_spans` {% /dt %} {% dd %} Load children of specific spans for progressive tree exploration when `get_llmobs_trace` returns collapsed nodes. {% /dd %} {% dt %} `get_llmobs_agent_loop` {% /dt %} {% dd %} Get a chronological view of an agent's execution loop, showing each step (LLM calls, tool invocations, decisions) in order. {% /dd %} {% /dl %} ### Experiment tools{% #experiment-tools %} {% dl %} {% dt %} `get_llmobs_experiment_summary` {% /dt %} {% dd %} Get a high-level experiment summary with pre-computed statistics for all evaluation metrics. Start here before using other experiment tools. {% /dd %} {% dt %} `list_llmobs_experiment_events` {% /dt %} {% dd %} List experiment events with filtering by dimension or metric and sorting by metric value. {% /dd %} {% dt %} `get_llmobs_experiment_event` {% /dt %} {% dd %} Get full details for a single experiment event, including input, output, expected output, all metrics, and dimensions. {% /dd %} {% dt %} `get_llmobs_experiment_metric_values` {% /dt %} {% dd %} Get statistical analysis for a specific evaluation metric, optionally segmented by a dimension for comparison. {% /dd %} {% dt %} `get_llmobs_experiment_dimension_values` {% /dt %} {% dd %} Get unique values for a dimension with counts, useful for discovering valid filter and segment values. {% /dd %} {% /dl %} ### Evaluator tools{% #evaluator-tools %} {% dl %} {% dt %} `list_llmobs_evals` {% /dt %} {% dd %} List every LLM-judge evaluator configured across all ML applications. Returns each evaluator's name, ml_app, and enabled status. {% /dd %} {% dt %} `list_llmobs_evals_by_ml_app` {% /dt %} {% dd %} List all LLM-judge evaluators configured for a specific ML application. {% /dd %} {% dt %} `get_llmobs_evaluator` {% /dt %} {% dd %} Retrieve an LLM-judge evaluator configuration by name, including its target (ml_app, sampling, filter), LLM provider, and judge prompt template. {% /dd %} {% dt %} `create_or_update_llmobs_evaluator` {% /dt %} {% dd %} Create or update an LLM-judge evaluator configuration. Targets a specific ML application and optionally a filter or sampling percentage; the judge's model and prompt template define how it scores each span. {% /dd %} {% dt %} `delete_llmobs_evaluator` {% /dt %} {% dd %} Delete an LLM-judge evaluator configuration by name. {% /dd %} {% /dl %} ## Recommended workflows{% #recommended-workflows %} ### Trace analysis{% #trace-analysis %} 1. **Search**: Use `search_llmobs_spans` to find traces by ML app, status, span kind, or custom tags. 1. **Visualize**: Use `get_llmobs_trace` to see the full span hierarchy tree. 1. **Inspect**: Use `get_llmobs_span_details` to get metadata, timing, and evaluations for specific spans. 1. **Read content**: Use `get_llmobs_span_content` to retrieve the actual I/O, messages, or documents. 1. **Debug errors**: Use `find_llmobs_error_spans` to locate all errors in a trace with propagation context. 1. **Expand**: Use `expand_llmobs_spans` to load children of collapsed spans for deeper exploration. 1. **Agent review**: Use `get_llmobs_agent_loop` to see the step-by-step execution flow of an agent span. ### Experiment analysis{% #experiment-analysis %} 1. **Summarize**: Use `get_llmobs_experiment_summary` to get overall statistics and discover available metrics and dimensions. 1. **Browse events**: Use `list_llmobs_experiment_events` to find events of interest, filtering by dimension or sorting by metric. 1. **Inspect events**: Use `get_llmobs_experiment_event` to view full details for a specific event. 1. **Analyze metrics**: Use `get_llmobs_experiment_metric_values` to get percentile distributions, true/false rates, or compare across dimension segments. 1. **Discover dimensions**: Use `get_llmobs_experiment_dimension_values` to find valid filter and segment values. ## Example prompts{% #example-prompts %} After connecting, try prompts like: - Review error traces for my `customer-support-bot` app over the past week. Summarize the most common failure patterns, how often they occur, and recommend which ones to fix first. - Find traces where my agent's responses were flagged by evaluations as low quality. Look at the inputs and outputs, then suggest specific changes to my system prompt to improve response quality. - Look at recent agent traces for my app and find cases where the agent looped more than necessary. Analyze the decision-making at each step and suggest how to improve my tool descriptions to reduce unnecessary tool calls. - A user reported a bad response. Here's the trace ID: `trace-123`. Walk me through exactly what happened: what the user asked, what the agent did at each step, and where things went wrong. Suggest a code fix. - Analyze experiment `exp-456` and generate a markdown table of the worst-performing dimensions broken down by evaluation scores. Include any other relevant columns that help me understand where and why performance is degrading. - Compare experiment `exp-123` (baseline) against experiment `exp-456`. Summarize what improved, what regressed, and by how much. Give me a recommendation on whether the changes are worth shipping. - Summarize experiment `exp-456` and identify the top 5 lowest-scoring events. For each, show the input, output, and which evaluations failed. ## Combine with other Datadog tools{% #combine-with-other-datadog-tools %} The `core` toolset included in the setup URL gives your AI agent access to additional Datadog tools that pair naturally with LLM Observability analysis. ### Export analysis to Datadog Notebooks{% #export-analysis-to-datadog-notebooks %} The `core` toolset includes `create_datadog_notebook` and `edit_datadog_notebook`, which let your AI agent create [Datadog Notebooks](https://docs.datadoghq.com/notebooks.md) directly from analysis results. You can export findings from agent chats into a collaborative, shareable notebook that lives in Datadog alongside your traces and experiments. Try prompts like: - Analyze experiment `exp-456`, identify the worst-performing dimensions, and export a summary report to a Datadog Notebook with a breakdown by evaluation scores. - Review error traces for my `customer-support-bot` over the past week and create a Datadog Notebook with the findings, including common failure patterns and recommended fixes. For custom visualizations that go beyond standard Datadog widgets, like comparison charts or quadrant plots, Notebooks also render [Mermaid diagrams](https://docs.datadoghq.com/notebooks/guide/build_diagrams_with_mermaidjs.md) natively. Try prompts like: - Analyze experiment `exp-456`, compare the `accuracy` scores across each prompt version, and export the results to a Datadog Notebook that includes a Mermaid bar chart of the average score for each version. - Analyze experiment `exp-456` and export a Datadog Notebook that plots each prompt version on a Mermaid quadrant chart with `relevance` on one axis and `accuracy` on the other. Identify which versions are underperforming on both dimensions. ## Further reading{% #further-reading %} - [Datadog MCP Server](https://docs.datadoghq.com/bits_ai/mcp_server.md) - [Set up and use LLM Observability Experiments](https://docs.datadoghq.com/llm_observability/experiments.md) - [Monitor your application with LLM Observability](https://docs.datadoghq.com/llm_observability/monitoring.md) - [Analyze LLM Applications with Claude Code Skills](https://docs.datadoghq.com/llm_observability/guide/claude_code_skills.md)