nicholas-kloster/nuclide-atlas

Nuclide Atlas is a tool that draws a picture of the AI services a team is already running. It points at localhost, a URL, a Kubernetes namespace, or a YAML inventory and produces a graph showing language models, endpoints, deployments, retrieval pipelines, vector databases, tools, and agents, with the edges between them showing what is wired to what. Clicking any node opens a side panel with its configuration and a metrics snapshot. The README opens with the problem. Teams running modern AI tend to have stacks they cannot fully describe: someone stands up Ollama on a dev VM, a platform team deploys vLLM behind an internal load balancer, an agent framework wires several tools and RAG pipelines together, and no one outside the team can say which one of those touches customer data. Inventory tools assume you already know the inventory, and generic dashboards show metrics but not topology. Atlas tries to fill that gap by making the running stack the source of truth. There is a 30-second demo path that uses bin/atlas-demo to spawn mock Ollama, vLLM, and Qdrant services on loopback and open a browser. For a real stack you run bin/atlas-bootstrap, which checks environment variables such as OPENAI_API_BASE and OLLAMA_HOST, sweeps a catalog of known LLM ports on localhost, reads docker ps for containers running known LLM images, optionally reads a Kubernetes namespace through kubectl, writes the result to config/atlas.yaml, then boots the stack with docker compose. Exit codes are stable so CI tools can branch on them. Multi-arch Docker images are published to GitHub Container Registry for amd64 and arm64. The map supports several actions. Clicking an agent and hitting Highlight Path dims everything outside its blast radius. A Trace Query animation walks a request through safety, RAG, tools, model, and back. A Live Pulse toggle re-probes endpoints every 15 seconds and colors them green, amber, or red. Risk badges dot nodes that fail deterministic rules, such as an internal API with no auth. There is also a search box and layer filters. Atlas is described as internal observability, not a scanner. It only contacts hosts you put in the config. For external discovery the author points at a separate tool called aimap, whose JSON output Atlas can read. It is MIT licensed and the repository has zero stars at the time of writing.