← back to the tree

Atlas

Self-hosted, source-configurable engineering platform for gen-AI, ML, and data work.

30+ integrated servicesone Compose stack2,800+ bootstrap tests

git clone https://github.com/thekaveh/atlas && cd atlas && ./start.sh

Atlas rendered in blue wireframe, holding aloft a glowing golden wireframe globe — the platform bearing its world of services.
Atlas bearing its world of services. Digital painting, 2026 — from the repository's brand set.

Standing up a real environment for gen-AI, ML, and data work usually means wiring thirty-odd services together by hand, then doing it again on the next machine. Atlas replaces that with a self-hosted platform you bring up from a single Docker Compose stack — and the point is not the container count, it is that the services arrive already wired to one another. Every consumer of a model reads one LiteLLM URL and key, Kong’s routes are generated from whichever services are active, and the monitoring layer comes with its scrape jobs and dashboards already pointed at the right places.

The roster is worth naming. The LLM core is LiteLLM over Ollama and the cloud providers (OpenAI, Anthropic, OpenRouter), with a TEI cross-encoder reranker beside it. The data tier runs the Supabase Postgres family — database, auth, storage, realtime, Studio — plus Weaviate with a CLIP module for vectors, Neo4j for graphs, Redis, and MinIO for S3-compatible object storage, and extends into a lakehouse of Apache Iceberg, Trino, Spark, and Zeppelin with Redpanda for Kafka-compatible streaming. Media and ingestion cover ComfyUI for image generation; Speaches, Parakeet, whisper.cpp, and Chatterbox for speech in both directions; Docling and Apache Tika for documents; Crawl4AI and SearxNG for the web. Agents and workflows bring n8n, Airflow with Celery and Flower, the Hermes and OpenClaw agent runtimes, LightRAG for graph-augmented retrieval, and a curated MCP server set. On top sit the apps: Open WebUI for chat, JupyterHub, MLflow, Label Studio, Jenkins, Verba, the Neo4j LLM Graph Builder, and a local deep-research agent — with Ray underneath for distributed Python and Prometheus, Grafana, Langfuse, Loki, and Tempo watching the whole thing.

None of it is monolithic. Every service is independently switchable between container, localhost, or disabled, and a chosen track — RAG, agentic engineering, creative, ML, data, or trading — sets sensible defaults for the discipline at hand. The same stack can back other projects as shared infrastructure, running standalone on a shared Docker network or vendored in as a Git submodule.

Highlights

  • Pre-integrated, not co-located — Zeppelin’s Spark interpreter arrives pointed at the in-stack Spark master, MinIO, and the Iceberg catalog; Trino queries the same lakehouse; LightRAG defaults its backends to Supabase, Neo4j, and Redis; Airflow ships LLM operators already wired to LiteLLM
  • One gateway for every model — LiteLLM gives the entire stack a single OpenAI-compatible URL and key, whether a model runs in local Ollama or at a cloud provider; even the Hermes agent surfaces as just another model to every consumer
  • Kong routes generated, not maintained — the bootstrapper derives gateway routes and the port map from whichever services are active, and a generated dashboard lists them all with live reachability checks
  • Interactive setup wizard — a Textual TUI that walks service configuration step by step, with dependency validation, a live command preview, and color-coded log streaming once the stack launches
  • Full data layer on first launch — Supabase Postgres, Weaviate, Neo4j, Redis, and MinIO come up ready, with the Iceberg–Trino–Spark lakehouse and Redpanda streaming a flag away
  • Observability that is already scraping — the opt-in Prometheus bundle ships fifteen scrape jobs across Kong, LiteLLM, Weaviate, n8n, MinIO, Postgres, Redis, and the backend, and Grafana pre-provisions seven starter dashboards
  • Shared infrastructure for other projects — designed to back downstream work over a shared Docker network or as a Git submodule; RAG Showcase runs on Atlas vendored exactly that way

Stack & Architecture

A Python bootstrapper reads the SOURCE variables, generates Kong’s routes and the port map, and drives the wizard. Each of the fifty-odd services is self-contained in its own services/<name>/ folder — manifest, Compose fragment, init scripts, config — and a thin top-level file composes them into one stack, so a new service is added in one place without touching the others. Each manifest also declares its data flows, and per-service architecture diagrams are regenerated from those declarations, so the integration map is derived from the same source of truth the stack runs on. The bootstrapper carries 2,800+ tests, and the platform runs natively on Linux and macOS, Intel or Apple Silicon, and on Windows through WSL2.

Status

Actively developed and Apache-2.0 licensed, with a full documentation site built from the in-repo docs.