Preserve the reason behind the files
KLYPIX is a portable project workspace with a shared brain. It keeps files, evidence, decisions, corrections, and current context together for people and supported AI tools.
A repository records code changes, but the reason a team rejected an approach may remain in a conversation that the next session cannot see. KLYPIX stores captured decisions, corrections, evidence and open questions in a portable project record. For software projects, brain.klypix sits beside the code and can travel through Git.
The desktop application presents that record visually. The open-source klypix-mcp engine makes it available to supported coding tools through Model Context Protocol. The project starts with only the understanding that people or agents actually record; it does not reconstruct an unwritten history.
A concrete workflow: record, retrieve, correct
The published walkthrough starts with a real decision and its evidence. A supported session records it, and a later session retrieves the project context. When the decision changes, an explicit correction supersedes the earlier record while retaining its history. The reviewer can then ask again and inspect the returned sources.
This provides a practical acceptance check for shared project knowledge: does the next session retrieve the current decision and its evidence? The website includes a product walkthrough, while the public repository supplies a conformance command for coordination behavior using isolated MCP clients. That check is not a retrieval-quality benchmark.
Retrieve evidence and keep its limits visible
The brain engine uses local lexical retrieval by default. An optional on-device embedding model enables semantic ranking, with lexical matching retained for identifiers and paths. A fresh installation does not install that model automatically. MCP returns selected context to the connected tool; that tool's model provider can receive what its agent reads.
Evidence anchors can connect a decision to the code it describes. A changed source fingerprint is a reason to recheck the decision, not proof that it is wrong. This distinction matters when an AI assistant uses older project knowledge to explain or modify current software.
Make overlapping work visible
A session can declare its task and the exact files it expects to change. Other connected sessions on the same machine can receive presence information and overlap warnings. These warnings are advisory: they do not prevent edits, compare the agents' conclusions or guarantee that another session read a message.
Integration depth varies by host. Some use lifecycle hooks; others rely on explicit MCP calls and project instructions. Durable decisions travel when the brain file is shared through the repository. In the standalone MCP package, live presence and overlap warnings are machine-local. The current source also documents an optional cross-machine presence core that requires desktop transport and consent; sharing the brain file alone does not enable it.
What this demonstrates for a custom AI project
KLYPIX demonstrates engineering around internal knowledge, source references, correction lifecycles and AI-tool integration. These are relevant when a business wants an assistant to use current decisions and explain where an answer came from. They are evidence of product capabilities, not measured savings or an existing deployment in a particular business system.
A custom implementation still needs its own source inventory, access rules, representative questions and acceptance criteria. KLYPIX is in beta; its desktop app is proprietary while the MCP engine is open source. Planned enterprise capabilities should not be treated as delivered features or as compliance guarantees.
Your next project
Discuss AI connected to your project knowledge
Explore the bilingual assistant architecture and runnable example