djovaneli/llmxray repository overview

⁠LLMxRay

If LLMxRay helps you understand your AI better, consider giving it a star on GitHub⁠.

Local LLM Observatory — See what your AI is actually doing, token by token, layer by layer.

No cloud. No API keys. No cost. Everything runs on your hardware.

⁠Quick Start

docker run -p 5174:5174 -p 11435:11435 djovaneli/llmxray

Open http://localhost:5174⁠ in your browser. Make sure Ollama⁠ is running on your machine.

The second port is the Live page's observe port: point any other tool (VS Code Copilot, Open WebUI, Continue, n8n, your own scripts) at http://localhost:11435 instead of Ollama and its traffic shows up on the Live page. Leave the mapping out if you do not need it. A published port is reachable from other machines on your network, so the Live page can run full-screen on a second screen while the agents work; bind it to 127.0.0.1:5174:5174 if you do not want that.

An agent that sends its calls through 11435 can also ask LLMxRay what went wrong with them: the image ships llmxray-mcp, a read-only MCP server (docker exec -i <container> node bin/mcp.mjs, or npx -y -p llmxray llmxray-mcp on the host with LLMXRAY_URL pointing at this container). See "For agents" in the main README.

⁠Configuration

⁠Sending turns to your trace viewer

docker run -p 5174:5174 -p 11435:11435 -e OTEL_EXPORTER_OTLP_ENDPOINT=http://host.docker.internal:4318 djovaneli/llmxray

Each model call becomes a span and each run a trace. The prompt text is never sent unless you switch it on in the Live page. docker stop lets the open run's spans out before the container ends.

⁠Keeping the Live log across restarts

Each turn's record is written under /data, and LLMxRay puts the last 500 back when it starts. A container keeps nothing between runs unless that folder is a volume:

docker run -p 5174:5174 -p 11435:11435 -v llmxray-data:/data djovaneli/llmxray

⁠Custom Ollama URL

docker run -p 5174:5174 -e OLLAMA_URL=http://192.168.1.50:11434 djovaneli/llmxray

⁠Custom port

docker run -p 3000:3000 -e PORT=3000 djovaneli/llmxray

⁠Docker Compose

⁠Standalone (Ollama running on your host)

services:
  llmxray:
    image: djovaneli/llmxray
    ports:
      - "5174:5174"
      - "11435:11435"
    environment:
      - OLLAMA_URL=http://host.docker.internal:11434
    volumes:
      - llmxray-data:/data

volumes:
  llmxray-data:

⁠With Ollama as a sibling container

services:
  ollama:
    image: ollama/ollama
    ports:
      - "11434:11434"
    volumes:
      - ollama-data:/root/.ollama

  llmxray:
    image: djovaneli/llmxray
    ports:
      - "5174:5174"
      - "11435:11435"
    environment:
      - OLLAMA_URL=http://ollama:11434
    depends_on:
      - ollama

volumes:
  ollama-data:

⁠Linux Note

On Linux, host.docker.internal may not resolve by default. Either:

• Use the Docker Compose example above with Ollama as a sibling container
• Add --add-host=host.docker.internal:host-gateway to your docker run command

⁠What you get

• Prompt X-ray: each turn's prompt drawn to scale, showing what was already cached, what was read again, and what Ollama silently cut when the prompt outgrew the context: the oldest messages, or the start of a system prompt too long on its own
• Runs: one tool's calls to the model, end to end, as a waterfall: waiting, loading, reading the prompt and writing the answer for each call, tool time between calls, and what every call sent and got back
• Shared box: why Ollama is slow right now and which tool did it, in seconds, with the setting to change (a keep_alive: 0 that unloads the model after every run, a 128K context on 4K prompts)
• GPU map: what is in video memory, model by model, and which pulled models would fit
• Tool calls: thirteen request shapes agents send, scored per model on what the daemon returned
• Live: what your other tools send to Ollama, turn by turn, and where their prompts break the cache
• Chat rendered as Markdown while it streams, every token coloured by confidence
• Cache Lab: find the value that defeats the prompt cache and measure the fix on your own daemon
• Compare: up to four slots side by side, with a word-level diff
• Surgical Benchmark on real token logprobs
• Decision Lab for Ollama's decision models, probabilities drawn as bars
• Protocol Observatory: native, OpenAI-compatible and Anthropic-compatible paths side by side, reasoning included
• Knowledge Base and embeddings, all in the browser's IndexedDB
• Code Completion, Tool Workshop, costs, analytics, system monitor
• Five languages, dark and light mode

⁠Links

• Website: lognebudo.github.io/llmxray⁠
• Documentation: lognebudo.github.io/llmxray/docs/en/⁠
• GitHub: github.com/LogneBudo/llmxray⁠
• npm: npmjs.com/package/llmxray⁠

⁠License

Apache 2.0 — Copyright 2026 Ivan Stankovic (LogneBudo⁠)