What this thing is
The Jetson Local AI API is my home-built HTTP service for giving machines on a trusted local network access to the AI capabilities running on the Jetson. It turns the board into a shared local AI bench: text, vision, image, and audio jobs for everyone on the LAN, free AI in the very literal sense that no cloud meter spins up every time an agent asks for a picture.
It sits in front of a ComfyUI workspace and exposes a friendlier operational surface for agents: health checks, model inventory, capability status, job creation, polling, generated assets, usage totals, and history.
The API listens on port 8899. From the box itself you use http://127.0.0.1:8899. From another local machine you use the LAN address. The documentation site lives under /site/, and the machine-readable map lives at /openapi.json.
Mental model: this is not a public SaaS product. It is a local workbench with an HTTP handle. Treat it like a power tool, not like a vending machine you found in the cloud and decided to shake.
The short story
A coding agent wants a picture. A frontier image API would take the prompt, bill the account, and return something suspiciously glossy. The Jetson API asks a ruder but better question: "Is the local box alive, is ComfyUI reachable, and do I have enough memory to survive this request?"
That is the useful part. The API makes local generation boring enough for agents to use. It wraps the tedious bits: queueing, polling, route discovery, assets, history, usage, and the humility required when an 8 GB edge board is asked to impersonate a data center. No, the board is not a cloud GPU wearing a tiny hat.
The useful shape
Inspect
Use /health, /capabilities, /models, and /workflows before asking for expensive work.
Create
Submit long-running work to /jobs/text, /jobs/vision, /jobs/image, /jobs/audio, or /jobs/workflow.
Poll
Read /jobs/{job_id} until the job reaches succeeded, failed, stopped, or cleared.
Fetch
Generated job files come back under /assets/{filename}. Benchmark artifacts have their own retained asset route.
First poke
From the Jetson itself:
API=http://127.0.0.1:8899
curl -sS "$API/health"
curl -sS "$API/capabilities"
curl -sS "$API/openapi.json"From another trusted LAN device, swap in the LAN base URL. If /health does not report API status ok and ComfyUI reachable, stop. Debug the machine. Do not keep flinging jobs at a dead backend like a racquetball player in a server closet.
What it can do today
The verified domains are local Qwen3 text generation, Qwen2-VL vision analysis, several image generation and replay paths, SD1.5 CyberRealistic img2img reference diffusion, a fast local reference stylize fallback, and Stable Audio Open short audio. The API also tracks local model inventory and retained benchmark evidence.
Some things are explicitly not ready. Video generation is unsupported for this Jetson profile. Hunyuan3D and BiRefNet background removal are documented-only in the handoff package. ControlNet Canny is a recommended optional upgrade, and IP-Adapter is intentionally not the first target on an 8 GB board because reality continues to have opinions.
Why this matters
The point is not to win a benchmark screenshot contest. The point is to make small, local, repeatable AI operations available to agents without forcing every experiment through a cloud toll booth. Some work belongs near the bench: reference edits, route smoke tests, local model checks, generated sprites, short audio, and "does this even work?" loops.
That thread connects directly to Learning by Osmosis on the Jetson Nano and the robot build note with a Jetson-class system2 brain. Same bench. Different flavor of cable management.