Why portable agents

The thesis

Local LLMs are about to become as personal as a phone. The dominant shape of an AI agent today is a tab on someone else's website — ChatGPT, Claude, Cursor, Replit. The owner of the agent and the owner of the chat surface are different parties.

GemmaPod is the bet that the next shape is different: the agent is the artefact. You write its persona, sign it, and send it.

This is the same trajectory email took. RFC 822 didn't promise a chat platform; it promised "you can write something and route it through servers you don't run to a person you do trust." A signed pod is the same promise applied to an AI agent: write it, sign it, route it.

Three things that change once the agent is portable

1. Who runs the model is decoupled from who designed the agent

You can ship a persona that runs on your Mac mini at home, on a friend's server, on a sidecar in a corporate VPC, or in the visitor's own browser via WebGPU. The persona is the same; the inference is wherever the trust is.

2. The system prompt is signed, not whispered

Today's chatbots have a server-side system prompt nobody outside the operator can see. A GemmaPod's system prompt is committed to a public signature. If the operator changes the persona, the signature changes, the pod id changes, and the change is visible. This matters for business-card use cases ("does this AI agent really represent Raj?") and for negotiation use cases ("did the persona I signed even reach the other party?").

3. Tools become a portable allow-list

A pod carries a signed list of tool names. The Host executes only tools that are both in the signed manifest and implemented in the local registry. The pod cannot grow new tools at runtime; the runtime cannot expose tools the pod didn't authorise. Trust is bounded to the union of two static sets, both controlled by the owner.

Three use cases the SDK was built for

The portable AI business card. Drop it on someone's website, attach it to a job-search email. It answers questions about you, shares a CV on request, runs on your home machine so you control the persona.
The take-home product demo. A SaaS company emails a prospect a .html that explains their product, answers questions, and lets the prospect kick the tires without a sales call.
The negotiator that goes places for you. Pods can carry the owner's persona + a tools allow-list through systems the owner never visits. (See the negotiation recipe.)

What GemmaPod deliberately doesn't try to do

Replace ChatGPT or Claude. Those are excellent at being chatbots. GemmaPod is excellent at being sendable.
Be a model platform. Inference happens via Ollama or transformers.js + WebGPU. The SDK is the envelope and the courier.
Have a default centralised cloud. The reference @gemmapod/signal signaling broker is 30 lines you run on your own VM. The public gemmapod.com/<id> deploy path exists for convenience, not lock-in.

If that landed, the next pages are practical: