GemmaPod

Composable, portable AI agent capsules — built on Gemma 4, signed end-to-end, and runnable anywhere a browser is.

A GemmaPod is an AI agent treated as a thing — a single signed .html capsule (~960 KB) carrying its identity, persona, tools, and transport. Forge it once, then ship it: email it, drop it into a web page with one <script> tag, or host it on your own URL.

Pods are objects that live on your device, not in someone's cloud. When a visitor opens one, it reaches back over WebRTC to a model running on your machine. If you're offline, it falls back to a small Gemma 4 model running entirely in the visitor's own browser via WebGPU. No middle server. No proxy. No chat plaintext touching the cloud.

Composable. Transportable. Embeddable. Headless when you want it, batteries-included when you don't.

Six things to know

   POD          one signed .html capsule  =  one agent

   ├─ TOOLKIT   build & sign:   pod.toml + owner.key → agent.html
   ├─ HOST      device runtime: registers pods, routes requests, owns memory
   │   ├─ MODEL   local LLM bridge (Ollama / OpenAI-compatible)
   │   └─ BUS     pod registry, event stream, conversation store
   ├─ SIGNAL    WebRTC handshake relay (no chat plaintext crosses it)
   ├─ EMBED     browser runtime: loads a pod inside any web page or email
   └─ gemmapod  the unified CLI — create, run, status, list, stop

You forge a Pod with the Toolkit, run it on your Host, expose it through Signal, and let other people Embed it. Same agent, many surfaces.

Why portable agents

Most "Gemma demo" projects end up as chatbots — useful, but the model is the whole product. GemmaPod inverts that: the model is the cheapest part of the system. The agent is the persona, the tools, the trust boundary, the conversation, the way it shows up on a page. We built the envelope around the model — a signed, transport-aware capsule you can actually hand to other people.

Examples of what a pod can be:

A portable AI business card that runs your persona on your home machine, embeddable on anyone's site.
A take-home product demo — a SaaS company emails a prospect a .html that explains the product and lets them kick the tires.
A restaurant or front-desk pod that takes orders and answers FAQs with a real cart driven by STATE_SNAPSHOT events.
A negotiator that goes places for you, carrying a signed allow-list of tools it's allowed to call.

Five-minute tour

What is GemmaPod?

Architecture & components

Build your first pod

Embed via a script tag

The Gemma 4 lineup, used intentionally

Variant	Where it runs	Why
Gemma 4 E2B	Visitor's browser via transformers.js + WebGPU	~1 GB at q4. The fallback that keeps the pod alive when the Host is offline.
Gemma 4 E4B	Owner's Host via Ollama	Primary path. Strong instruction following over a WebRTC data channel.
Gemma 4 31B Dense	Same Ollama endpoint, opt-in per request	"Heavy mode" for pods that need deeper reasoning.
Gemma 4 26B MoE	Optional sidecar	High-throughput batch tools (RAG, planning).

Project status

GemmaPod ships seven npm packages under the @gemmapod scope, plus Docker images for the Host and the signaling broker. v0.1 freezes the public surface; minor bumps may still tweak APIs as the SDK settles.

Source on GitHub · Changelog · Security model · MIT licensed