tempo-llm-cli

# Tempo LLM CLI LLM Agent designed to interact with Claude and the Tempo MCP Server. Use natural language to ask questions about your tracing data in Tempo. This is primarily a demo project to investigate the creation of LLM agents and interactions with MCP servers. However, feel free to use, extend, and PR improvements if you find this work interesting. ## Quick Start 1. `docker-compose up` the [local example](https://github.com/grafana/tempo/tree/main/example/docker-compose/local) in the Tempo main repo 1. If [this PR](https://github.com/grafana/tempo/pull/5212) is not merged you'll need to use the PR branch. 1. This starts Tempo exposed at http://localhost:3200 and pushes synthetic data into it. 1. Set your claude API ```bash # Set your Claude API key export ANTHROPIC_API_KEY=your_key_here ``` 1. Go! ``` # Build and run go run ./ ``` Now you can ask Claude things like: - "What's broken right now?" - "Show me the slowest traces from the last hour" - "Which service is causing errors?" - "Draw me a chart of response times" ## What You Need - Go 1.24+ - Claude API key from Anthropic ## Sample Chat ``` 👱 You: which endpoint is slowest right now? 🤖 Assistant: I'll help you find the slowest endpoint. Let me first check what attributes are available for endpoints and then query for the slowest one using metrics. 🔧 Tool Use: get-attribute-names map[scope:span] 📊 Tool Result (🏷️ Attribute Names): span (47 attributes): article.count http.flavor http.method http.request.header.accept http.request_content_length http.response.header.content-type http.response_content_length http.scheme http.status_code http.target http.url k6.0S5X7D9SjTA6B4N k6.1ciqVF24QgtjigU k6.1th67BzXcPKTngn k6.38qEKgBDP0EmRAz k6.3uHnIdM5hNzIEoW k6.4WHhMaOIX3uQyXi k6.8EsxHPjYAoFoq84 k6.9BQAWXEWC4WygVU k6.A0SUinRkUc8Exza ... (truncated after 20 attributes) 🤖 Assistant: Let me check for endpoint-related attributes and see what endpoint values are available: 🔧 Tool Use: get-attribute-values map[name:span.http.target] 📊 Tool Result (🏷️ Attribute Values): string values (10): /place-articles /set-billing /delete-product /update-payment /list-articles /authenticate /set-stock /payment /article-to-cart /get-article 🤖 Assistant: Now let me get the TraceQL metrics documentation to understand how to query for the slowest endpoint: 🔧 Tool Use: docs-traceql-metrics map[] 📊 Tool Result (📚 Documentation): TraceQL supports `rate`, `count_over_time`, `sum_over_time`, `min_over_time`, `avg_over_time`, `quantile_over_time`, `histogra... 🤖 Assistant: Now I'll query for the average duration by endpoint to find the slowest one: 🔧 Tool Use: traceql-metrics-instant map[query:{ span.http.target != nil } | avg_over_time(duration) by (span.http.target)] 📊 Tool Result (📊 Metrics Instant): Metric{span.http.target=/article-to-cart} : 0.7930 Metric{span.http.target=/authenticate} : 0.1995 Metric{span.http.target=/create-shipping} : 0.6470 Metric{span.http.target=/delete-product} : 0.6538 Metric{span.http.target=/get-article} : 0.3949 Metric{span.http.target=/list-articles} : 0.4300 Metric{span.http.target=/payment} : 0.1577 Metric{span.http.target=/place-articles} : 0.3846 Metric{span.http.target=/send-payment} : 0.6581 Metric{span.http.target=/set-billing} : 0.6455 ... (showing first 10 of 12 series) Search Metrics: Inspected Traces: 0 Inspected Bytes: 406686 Inspected Spans: 2589 Total Blocks: 2 Completed Jobs: 3/2 Total Block Bytes: 1562937 🤖 Assistant: Based on the metrics results, **`/article-to-cart`** is currently the slowest endpoint with an average duration of **0.793 seconds** (793 milliseconds). Here's the ranking of endpoints by average response time: 1. **`/article-to-cart`** - 793ms (slowest) 2. **`/set-stock`** - 692ms 3. **`/update-payment`** - 676ms 4. **`/send-payment`** - 658ms 5. **`/delete-product`** - 654ms 6. **`/set-billing`** - 646ms 7. **`/create-shipping`** - 647ms 8. **`/list-articles`** - 430ms 9. **`/get-article`** - 395ms 10. **`/place-articles`** - 385ms 11. **`/authenticate`** - 200ms 12. **`/payment`** - 158ms (fastest) The `/article-to-cart` endpoint is significantly slower than most others, taking nearly 800ms on average compared to the fastest endpoint (`/payment`) at 158ms. ```