The route groups
The Jetson API is deliberately boring in the best possible way. It splits the world into inspection routes, documentation routes, assets, history, and job routes. Agents can inspect before acting, create work as jobs, poll until completion, and fetch outputs without spelunking through the ComfyUI file tree like a confused intern.
| Group | Routes | What to use it for |
|---|---|---|
| Health and inventory | /health, /capabilities, /models, /workflows |
Find out what is alive, installed, verified, and available before you start expensive work. |
| Evidence | /benchmarks, /benchmark-summary, /benchmark-assets/{filename} |
Read retained benchmark results and media. This is the "prove it" drawer. |
| Job assets | /assets/{filename} |
Serve uploaded reference copies and generated API outputs. |
| Operations | /throttle, /usage, /history |
Check queue pressure, token-like accounting, and previous requests. |
| Jobs | /jobs/text, /jobs/vision, /jobs/image, /jobs/audio, /jobs/workflow |
Create work, then poll /jobs/{job_id} and read outputs. |
The capability matrix is the adult in the room
GET /capabilities is where the API stops being aspirational. Each item has an id, modality, status, route, limits, evidence, and throttle class. If a capability is verified, it has a route and supporting evidence. If it is documented_only, it may exist in the model folder but has not passed the API smoke bar. If it is unsupported, believe it. The machine is not being shy.
API=http://127.0.0.1:8899
curl -sS "$API/capabilities"
The current verified set includes local Qwen3 1.7B text, Qwen2-VL vision, Segmind-Vega image replay, CyberRealistic V9 image paths, SD1.5 CyberRealistic img2img, local reference stylize, and Stable Audio Open short audio.
Verified is not documented-only
This distinction saves time. Video generation has downloaded-looking things in the ecosystem but is unsupported for this Jetson profile. Hunyuan3D and BiRefNet background removal are documented-only. ControlNet Canny is a reasonable optional upgrade. IP-Adapter is not installed and is not the first target because the board has 8 GB, not a portal to a secret GPU bunker.
Agent rule: read /capabilities before planning work. If the thing is documented-only, write like an engineer: "not verified yet." Do not rename hope as architecture.
Models are inventory, not promises
/models scans local model directories and reports names, paths, modality hints, formats, sizes, and basic hashes. This is useful for discovery, but model presence is not enough. A seven-gigabyte checkpoint sitting on disk is not a working product. It is a large file with dreams.
curl -sS "$API/models"
The important installed highlights include Qwen3-1.7B-Q4_K_M.gguf, Qwen2-VL-2B-Instruct-Q4_K_M.gguf, CyberRealistic_V9_FP16.safetensors, segmind-vega.safetensors, stable-audio-open-1.0.safetensors, and the Z-Image-Turbo plus Flux2-related files.
The route-table story
The happy path is not "know the codebase." The happy path is "ask the API." An agent should be able to discover the route table at /openapi.json, inspect human docs at /docs or /site/api, then choose a verified job route. That is what makes the system usable by agents that did not write it yesterday and forget what they meant by "temporary." We have all met yesterday's temporary.