Open Source Router – Auxot Docs

What is the Auxot Router?

The Auxot Router is a stateless, open-source inference router that exposes OpenAI- and Anthropic-compatible APIs and routes requests to GPU workers, CLI workers, or tools workers. It is a single Go binary with no database — state lives in an embedded miniredis instance (or an external Redis if you’re scaling horizontally).

There are no agents, no skill libraries, no multi-user accounts, and no license keys. Authentication is a pair of Argon2id-hashed API keys: one for callers, one for workers.

┌───────────────┐     ┌───────────────┐     ┌───────────────┐
│ GPU Worker    │     │ CLI Worker    │     │ Tools Worker  │
│ (llama.cpp)   │     │ (Claude Code) │     │ (code,search) │
└───────┬───────┘     └───────┬───────┘     └───────┬───────┘
        │  WebSocket          │                     │
        └─────────┬───────────┘─────────────────────┘
                  │
          ┌───────▼───────┐
          │ auxot-router  │ ← OpenAI / Anthropic API
          │  (Go binary)  │
          └───────┬───────┘
                  │
          ┌───────▼───────┐
          │     Redis     │ ← Embedded (miniredis)
          │ (or external) │    or external for scaling
          └───────────────┘

Key Characteristics

No database — zero Postgres, zero migrations, zero schema management
No external services required — embedded miniredis works out of the box for single-instance deploys
Single binary — auxot-router is a ~10 MB static binary. The Docker image uses FROM scratch with no OS layer
700+ model registry — GGUF model catalog compiled into the binary, auto-downloaded from HuggingFace
Three worker types — GPU (llama.cpp), CLI (Claude Code), Tools (code execution, web search, MCP)
Fully open source — Apache 2.0 license, source at github.com/auxothq/auxot

Three Binaries

Binary	Role
`auxot-router`	The router process. Accepts API requests, dispatches jobs to workers via WebSocket
`auxot-worker`	GPU or CLI worker. Connects to the router, runs inference, streams tokens back
`auxot-tools`	Tools worker. Executes tool calls (code, search, MCP) on behalf of the LLM

When to Use the Commercial Server Instead

The open-source router is the right choice for individuals and small teams who want self-hosted GPU inference with minimal setup. Note that Auxot Server uses port 8420 by default (vs 8080 for the OSS router). If you need any of the following, see Auxot Server →:

Multi-user accounts, teams, and RBAC
Persistent agents with configurable skills and context files
MCP server governance and per-agent tool policies
License-gated features and enterprise support