Three questions, in order: what you need, which one you choose, and your first reuse working in about 10 minutes. Everything on this page is copy-paste.
| Item | Requirement |
|---|---|
| A machine | 64-bit Linux, ARM64 (a Jetson Orin, another ARM64 board) or x86_64 (a workstation, server, or cloud instance). The easy path: Ubuntu 24.04 for x86_64, Ubuntu 22.04 / JetPack for ARM64. No GPU needed for NeuraFrame™ itself. |
| Python | The exact CPython your build was compiled for: x86_64 builds need 3.12 (Ubuntu 24.04's default), ARM64/Jetson builds need 3.10 (Ubuntu 22.04's default). The preflight checks and tells you plainly rather than half-installing. |
| Memory | 8 GB RAM minimum (the preflight enforces it; the floor leaves room for a model running beside NeuraFrame™, so fronting only a hosted provider it is conservative). |
| Root | sudo for the install and for licensing commands (they write a license file). |
| A model | Something for NeuraFrame™ to sit in front of: a hosted provider API (OpenAI, Anthropic, Grok, or compatible), a local model served over HTTP, or a vision model endpoint. Embodied's bundled example needs none. |
| Internet | Once, to download and to activate the 7-day trial (offline activation is available by request). |
Choose by what you are running, not by product name. One fact makes this simpler than it looks: Studio and the Gateway are the same download; you pick the mode during install and can change it later. Embodied and Fleet are their own packages.
| You are running… | Choose | Why |
|---|---|---|
| An app that calls a hosted provider (OpenAI, Anthropic, Grok, compatible) | Gateway (the Studio download, proxy mode) | One base-URL change in your app, zero code change. Repeats, paraphrases, and pinned answers served from memory; everything else forwarded unchanged. |
| A local model on your own hardware | Studio (the same download, native mode) | Reuse plus teaching: ask and correct through the API, and corrections change future answers. |
| A vision model (classifier or detector), images or video | Studio + vision reuse | Reuse by the entities in the image, and on video follow a thing across frames instead of re-classifying it every frame. Vision reuse. |
| A robot or machine you want to train in simulation | Embodied | Raise the memory in sim, ship it as one file, no model on the robot. The bundled example runs in minutes; wiring your real robot is an integration project. Why Embodied. |
| Many devices that should share what they learn | Fleet (added later) | Start with one device today; Fleet adds curated round-trip learning when you scale. Studio Fleet. |
Both paths start the same way: download, install, start the trial. Grab the build for your CPU from the download page (uname -m: aarch64 means ARM64, x86_64 means x86_64).
# 1. install (about a minute; the preflight checks your system first) tar -xzf neuraframe-studio-server-0.6.1.tar.gz cd neuraframe-studio sudo bash install.sh # 2. start the 7-day trial (needs sudo: it writes a license file) sudo neuraframe trial start
# 3. point the gateway at your provider (auto profile detects OpenAI and Anthropic style) sudo neuraframe config set gateway.upstream_url https://api.openai.com sudo systemctl restart neuraframe-studio # 4. point your app at the gateway instead of the provider. That is the whole code change. # base URL: http://127.0.0.1:8081/v1 (Anthropic style: http://127.0.0.1:8081) # your API key travels on the request exactly as it does today # 5. send the same request twice, then look at what you saved neuraframe savings
# 3. tell NeuraFrame where your model listens sudo neuraframe config set completion_url http://127.0.0.1:8080/completion sudo systemctl restart neuraframe-studio # 4. ask twice: the first answer comes from your model, the repeat from memory neuraframe ask "What is our return policy?" neuraframe ask "What is our return policy?" # 5. teach it something and watch the answer change from then on neuraframe correct "the old answer" "the corrected answer"
neuraframe vision on
neuraframe vision mode near # or scene / continuous; see the vision guide
sudo systemctl restart neuraframe-studio
Point the gateway's upstream at your vision model and send frames as usual; the vision guide has every mode and control, including video.
neuraframe status # service healthy, licensed, engine active neuraframe savings # calls avoided, reuse rate, tokens saved, by source neuraframe doctor # if anything looks off, this says what and why
If savings shows calls avoided climbing, you are done: NeuraFrame™ is serving repeats from memory and your model is only doing new work. From here, the docs cover teaching, freshness, pins, vision modes, and fleets.
neuraframe doctor first: it checks your setup and tells you what is wrong. If that does not solve it, the troubleshooting guide covers the usual suspects, and contact us gets you a person, not a bot.