Gemini CLI + MCP Tools Deep Dive - Build a Completely Local RAG with Ollama | Context7, NextJS

Gemini CLI can be paired with an MCP server (Context7) to generate and run a fully local RAG-style “chat with your files” web app—complete with streamed model responses and library-aware scaffolding. The practical payoff is that Context7’s MCP tools pull up-to-date documentation for the chosen web stack (Next.js, Shadcn UI, Tailwind, and related libraries), letting Gemini CLI scaffold a working Next.js application and wire it to a model that answers questions grounded in uploaded text files.

The build starts with installing Gemini CLI via Homebrew (brew install Gemini CLI) and adding a Context7 MCP server connection into Gemini CLI’s configuration. Once configured, Gemini CLI uses Context7 tools such as “resolve library ID” and “get library docs” to ingest documentation for the target stack. In the author’s run, Gemini CLI scraped documentation for Tailwind and other libraries, then attempted to scaffold the Next.js app using create-next-app. A minor snag—an existing non-empty directory—was handled by generating the project in a subdirectory. The scaffolding phase also surfaced dependency and tooling issues, including Shadcn UI being flagged as deprecated and the need to switch to the recommended Shadcn package.

After the initial project structure was created, Gemini CLI focused on the UI and the core streaming behavior. The app adds a chat interface on the landing page, then implements server-side streaming so responses arrive chunk-by-chunk rather than waiting for a full completion. The transcript shows an async readable stream that iterates over response chunks and pushes them into a controller, with the client consuming the stream while rendering markdown.

For the “chat with files” part, the app includes an upload flow for text files. Uploaded content is passed into the chat component and injected into the prompt as a system message (“The user has provided the following document content”), while the conversation history is sent as the remaining messages. This setup is then used to generate answers grounded in the uploaded file. The UI renders assistant messages through react-markdown, and the build later switches models from Gemini 2.5 Pro to Gemini 2.5 Flash while keeping the same overall workflow.

Styling and layout required iterative fixes. Tailwind’s typography plugin caused overlapping text in the response bubble, and the bubble width and “thinking” UI placement needed adjustments. The final layout improved by narrowing controls and placing the “thinking” section above the rest of the message, while an accordion component supports collapsing and expanding the reasoning/summary area.

In runtime testing, the app produced a coherent summary of the uploaded file and continued the conversation by recalling that summary. A known UI bug remained: the page scrolls the entire website instead of only the chat area. Performance-wise, the transcript reports roughly 10 turns, about 11 million tokens total (as estimated), and an 11-minute API duration, leaving 88% of context unused at the end.

Overall, the workflow demonstrates a repeatable path: use Context7 MCP to keep a generated Next.js stack aligned with current library docs, then rely on Gemini CLI to wire streaming chat and file-grounded prompting into a local web app that can be extended into a more complete RAG system.