jmagar/axon

Axon

Version: 5.9.0

Axon is a self-hosted RAG stack for crawling, scraping, ingesting, embedding, searching, and asking questions over indexed content. The production release is Docker Compose first: one Axon server container, Qdrant, Hugging Face TEI with Qwen/Qwen3-Embedding-0.6B, and Chrome for JS-heavy pages.

Production Contract

Supported production runtime:

• Docker Compose only.
• Qdrant only for vector storage.
• Hugging Face TEI only for embeddings.
• Qwen/Qwen3-Embedding-0.6B as the production embedding model.
• Gemini CLI is the default LLM synthesis path; OpenAI-compatible endpoints such as llama.cpp are supported when configured with AXON_LLM_BACKEND=openai-compat.
• Web search/research uses a self-hosted SearXNG instance (AXON_SEARXNG_URL) when configured, falling back to Tavily (TAVILY_API_KEY) otherwise.
• Local NVIDIA RTX 4070 target with NVIDIA Container Toolkit.
• CLI and MCP run all actions in-process; deploy the axon serve container only when you need HTTP API access.
• One shared config home: ~/.axon/.env, ~/.axon/config.toml, ~/.axon/jobs.db, ~/.axon/output, ~/.axon/logs, ~/.axon/artifacts, ~/.axon/screenshots, ~/.axon/qdrant, and ~/.axon/tei.

Not supported in the production path:

• systemd deployment of the Axon binary.
• Postgres, Redis, RabbitMQ, AMQP, or external worker services.
• OpenAI-compatible first-run LLM configuration. Configure AXON_LLM_BACKEND=openai-compat manually after setup when using llama.cpp or another OpenAI-compatible /v1/chat/completions endpoint.
• Neo4j or graph retrieval.
• Multiple competing .env or config.toml locations.

Install

Linux

Prerequisites:

• Linux x86_64.
• Docker and Docker Compose.
• NVIDIA driver, nvidia-smi, and NVIDIA Container Toolkit.
• Gemini CLI installed and already authenticated, unless using a configured OpenAI-compatible endpoint for LLM synthesis.
• curl, sha256sum, and install.

One-line installer:

curl