I built GemmaPod - A truly composable and portable AI agent solution powered by your local LLM

What I Built

GemmaPod is a composable, portable AI agent platform that packages local Large Language Models into single, signed HTML+JS+WASM files (~960 KB).

A GemmaPod bundles an AI agent's complete identity, persona, tools, and transport mechanism into one self-contained, tamper-proof capsule.

It is composable — easily mix different tools, personas, and transports into a single signed file.

It is portable — you can email it, embed it with a single tag, or publish it to a stable URL.

And it speaks MCP and A2A out of the box, allowing your agent to connect to the tools and services it needs, while keeping the intelligence and your data private on your machine or your own deployments.

When activated, the pod connects back to home your local llm over DARTC — a real-time transport for portable AI agents. DARTC defines a signed, topic-multiplexed message envelope over WebRTC DataChannels (with a WebSocket relay fallback and first-class A2A alignment). It connects to your local LLM/Gemma models or runs entirely in the visitor's browser via WebGPU if the owner is offline.

*Here is the high level architecture

Main Links

• Website: gemmapod.com
• GitHub Repository: github.com/apprider/gemmapod

Demo

You can test GemmaPod instantly from your terminal:

npx gemmapod create

This command will create a ready-to-use Agent in your present working directory.
Requirements:

• Have Ollama running locally, OR
• Provide an OpenAI-compatible LLM URL (OpenAI, Groq, Together.ai, or any other compatible endpoint works perfectly for testing)

Once the pod is created and your LLM is running, simply open the generated signed HTML file in your browser. You can immediately start chatting with the agent.
The Agent comes with all the harness and workflow built-in. Vercel AI SDK is included so you can create complex workflows if needed. You can also couple the Mastra framework to build long sustained, multi-step workflows.

Code

Repository: github.com/apprider/gemmapod

Why I Built This

Local LLMs are incredibly powerful, but they lacked an effective and secure way to package and publish serious, multi-workflow AI agents as true consumer-grade products.

We needed a system that could bundle the complete agent — its harness, tools, workflows, persona, and transport — into one single, tamper-proof, cryptographically signed capsule.

While REST over HTTP revolutionized how we build and ship traditional applications, DARTC + GemmaPod brings that same level of simplicity, security, and portability to locally powered AI agents.

Live example: The intelligent agent that powers the dynamic Hero section on the GemmaPod website is actually running live on my Mac Mini right now.

Comments

[Sachin Kumar]

This is a very compelling evolution in how we think about AI agent deployment.
The idea of packaging the entire agent lifecycle (persona, tools, workflows, and transport) into a single, signed, portable unit is particularly powerful.
What stands out to me is:

• The local-first design, addressing data privacy and regulatory concerns
• The use of DARTC over WebRTC, which aligns well with low-latency, distributed agent interaction patterns
• The analogy to REST for applications—this feels like a similar inflection point for agent-based architectures

Overall, this is a strong step toward making AI agents truly portable, secure, and product-ready, especially for edge and enterprise use cases.