Local-First Agent Stack Demo

A new YouTube walkthrough shows developers wiring a local agent stack around Google’s Gemma 4, LM Studio, OpenCode and Paperclip on one machine. (youtube.com) The April 11, 2026 video from NetworkCoder says the setup runs “100% locally” and uses Gemma 4 with Qwen 3.5 2B inside a Paperclip multi-agent workflow. The clip had about 2,681 views and 77 likes when it was crawled. (youtube.com) The basic idea is to split the stack into layers. Gemma 4 is the model that generates text, LM Studio is the local engine that loads and serves the model, and OpenCode is the coding interface that sends requests and manages work inside a project. (blog.google ) (ai.google.dev) (opencode.ai) Google released Gemma 4 on January 15, 2026 in E2B, E4B, 31B and 26B A4B sizes. Google says the family is built for reasoning and agentic workflows, and its LM Studio integration supports local text generation, tool use and, in some variants, image understanding. (ai.google.dev 1) (ai.google.dev 2) LM Studio is the middle piece that makes a local model look like a service. Its documentation says it can serve requests through OpenAI-compatible endpoints, including chat, responses, completions and embeddings, which lets outside tools talk to a model running on the same computer. (lmstudio.ai) OpenCode sits on top of that service as the developer-facing layer. Its docs describe it as an open-source coding agent available in a terminal interface, desktop app or integrated development environment extension, and its provider docs say it can connect to local models through OpenAI-compatible back ends. (opencode.ai 1) (opencode.ai 2) Paperclip adds a fourth layer in the demo: orchestration across multiple agents. The project’s GitHub page describes it as a Node.js server and React user interface for assigning goals, tracking work and coordinating a team of agents from one dashboard. (github.com) That modular layout is the point of the demo. If one model underperforms, the video says developers can swap models in LM Studio without replacing the interface in OpenCode or the orchestration logic in Paperclip. (youtube.com) The pattern also lines up with how local artificial intelligence tooling is being packaged in 2026. Google is pushing Gemma 4 as an open model for on-device agents, LM Studio is exposing local models through standard application programming interfaces, and OpenCode is pitching model choice rather than a single-provider workflow. (developers.googleblog.com) (lmstudio.ai) (opencode.ai) OpenCode’s GitHub repository showed about 141,000 stars when it was crawled, and Paperclip showed about 52,200. The video’s appeal is that those pieces can now be combined into one local stack instead of being used as separate experiments. (github.com 1) (github.com 2)