Skip to content

Quickstart

A 2-minute tour. Prerequisites: installation done.

1. Start the agent

thor-cosmos
# 🤖🌌 thor-cosmos agent — ready
#     model = global.anthropic.claude-opus-4-6-v1
#     tools = 19
#     type 'exit' or Ctrl-C to quit
# 🌌 ▸

2. Talk to it

🌌 ▸ what's the state of the VLM server?

The agent calls cosmos_serve(action="status") → just serve-status → returns 🔴 not running.

🌌 ▸ download Cosmos-Reason2-2B and tell me what it is

The agent chains cosmos_model_download(name="reason2-2b") with a knowledge lookup.

3. Or use just directly

Every capability is a shell recipe — the agent and the operator share it:

just --list
#   default
#   env
#   install
#   run
#   deploy-thor
#   download
#   download-dataset
#   quantize
#   export-llm
#   export-visual
#   ...
#   smoke

4. Run a real pipeline

Prep a model on your x86 host

just prep-edge-model reason2-2b ./models/R2-fp8
# → download from HF
# → quantize to FP8
# → export LLM to ONNX
# → export visual encoder to ONNX

Deploy to Thor

just deploy-thor cagatay@thor.local ~/thor-cosmos
scp -r ./models/R2-fp8/onnx cagatay@thor.local:~/R2-fp8-onnx

On Thor: build engines + serve

ssh cagatay@thor.local
cd ~/thor-cosmos
just build-engines ~/R2-fp8-onnx ~/R2-fp8-engines
just serve-start ~/R2-fp8-engines/llm ~/R2-fp8-engines/visual
just serve-status      # 🟢 running pid=1234  http://127.0.0.1:8080

Inference

just infer assets/test.jpg "count people and describe their clothing"

Or through the agent:

🌌 ▸ capture a frame from RTP and tell me what you see

The agent calls rtp_capture_frame(...) → cosmos_inference(...) in one turn (the frame bytes are embedded in the first tool result, so the second tool sees the image directly).

Next