a2go（a2go 命令行）

a2go

Use open weight models (LLM, image, audio) with open source agents on Mac, Linux, and Windows.

Prerequisites

Requires the a2go CLI. Install from GitHub releases (includes SHA256 checksums for verification): https://github.com/runpod-labs/a2go/releases

Quick start

a2go doctor                                              # One-time setup (checks Docker, GPU, pulls image)
a2go run --agent hermes --llm <repo>:<bits>bit          # Start with a model
a2go status                                              # Check running services
a2go stop                                                # Stop all

Pick a model value with a2go models; use the repo:bits value from the output.

Commands

a2go run --agent <agent> --llm <repo>:<bits>bit [--image <repo>] [--audio <repo>:<bits>bit] [--engine <engine>]
a2go doctor                                              # Prereq check + image pull
a2go status                                              # Service health
a2go stop                                                # Stop containers

Agents: hermes (recommended) or openclaw.

Engines

• llama.cpp — NVIDIA GPU (CUDA). Default on Linux/Windows. Uses GGUF models.
• MLX — Apple Silicon. Default on Mac. Uses MLX models.
• wandler — ONNX runtime. Works on all platforms (CUDA on Linux, CPU fallback). Uses ONNX models. Pass --engine wandler to use it.

The engine is auto-detected from the model when possible. Use --engine to override (e.g. --engine wandler for ONNX models).

Ports

• 8000 — LLM API (direct model access, use for testing chat completions)
• 8080 — Web proxy / media server (TTS, STT, image gen, web UI)
• 8642 — Hermes Gateway (agent orchestration, not for direct API calls)
• 18789 — OpenClaw Gateway (agent orchestration, not for direct API calls)

For direct LLM testing use port 8000 (/v1/chat/completions). For TTS/STT use port 8080 (/v1/audio/speech, /v1/audio/transcriptions). The gateway ports (8642/18789) are for platform integrations.

Models

a2go models                                # All models
a2go models --type llm                     # LLMs only
a2go models --engine wandler               # Wandler/ONNX models only
a2go models --os mac                       # Mac/MLX models only
a2go models --max-vram 24                  # Fits in 24GB GPU
a2go models --type llm --engine wandler    # Wandler LLMs only

Output: type | engine | os | vram | context | repo:bits | name — use repo:bits as the --llm/--image/--audio value.

Docker

Image runpod/a2go:latest, configured via A2GO_CONFIG env var:

{"agent":"openclaw", "engine":"wandler", "llm":"onnx-community/gemma-4-E4B-it-ONNX:4bit"}

Fields:

• agent — hermes or openclaw (required)
• engine — llamacpp, mlx, or wandler
• llm — model as repo:bits from a2go models
• audio — audio model (optional)
• image — image model (optional)
• contextLength — override context window (optional)

Env vars: A2GO_AUTH_TOKEN (gateway auth), A2GO_API_KEY (LLM API auth).

Notes

• Mac/Apple Silicon: a2go run runs natively via MLX (no Docker). Wandler models also work on Mac.
• Browse models visually: https://a2go.run