Developer implements MCP and an agent loop in pure Java

On 2026-11-01, DEV Community's AI channel published a walkthrough titled "Building an MCP Client, Server, and Agent Loop in Java from Scratch." Its author built a working MCP server, an MCP client, and an agent loop entirely in plain Java and Spring Boot, deliberately avoiding any AI framework or MCP SDK. Instead, the system was wired to a Qwen 2.5 7B model served locally through Ollama, so all inference happened on the author's own machine rather than through a hosted API.

The whole piece is organized around a single, deliberately mundane request. The author asked the local model to add the phrase "buy milk" to a notes file, and the model did exactly that. What interests the author is not the outcome but the machinery behind it: the request had to be turned into a structured call, the intended operation had to be matched to a tool, the chosen tool had to be invoked with correctly formed arguments, and the result had to be fed back into the agent loop in a way that kept the conversation coherent. Having observed the behavior, the author then traced each step of that cycle to see precisely how the decision was made and where each component contributed. One note on method: the author is explicit that "from scratch" means refraining from frameworks, not from assistance. Claude co-authored both the code and the article, and the author remarks that hand-writing every line in 2026 would apparently be a pointless exercise. The article therefore doubles as a case study in how far a minimal, framework-free implementation can go when an agent is given a single, well-scoped task.

For now, the project remains a personal experiment: the article documents the setup and the traced call flow, but no repository, packaged implementation, or follow-up post has been published.