Klypix MCP provides shared project memory for coding agents. It is aimed at developers who use multiple coding tools at the same time (such as Claude Code, Codex, Cursor, etc.), solving the problem of context being lost between different tools and sessions.
Purpose
When you run multiple coding agents on the same codebase, context saved in one tool may be missing in another. Klypix MCP lets project decisions, corrections, and open questions carry across sessions and across tools through a brain.klypix file committed alongside the code. It does not start, run, supervise, or replace your agents, nor is it an agent runtime, model router, worktree manager, or a replacement for Git. It only stores what the project currently believes.
Main capabilities
- Decision lifecycle: A new decision that contradicts an old one supersedes the old one. Outdated cards are archived and marked with an arrow and date, not deleted; subsequent answers present the correction rather than the outdated content. If a later decision returns to an earlier superseded position, a high-confidence lineage leaves a dated re-adoption marker on the new card and preserves the original history.
- Explicit corrections: Supersession is triggered by an uppercase correction prompt or an explicit edge. brain_reconcile only proposes outdated-and-correction pairings for human confirmation.
- Evidence anchors: Cards can reference the code their decision is based on. ev: anchors record file:line number and the git blob OID at capture time, and the engine can flag cards whose referenced code has changed. It detects that the code has changed, rather than asserting that it has become false.
- Position meaning: Drag a card into the 📌 Focus area, and it will be shown with priority in the briefing of every future session. This is briefing priority, not a retrieval ranking boost.
- Historical queries: brain_ask with as_of: 2026-03-01 can reweight rankings by card lifecycle dates, so that corrections made later do not leak backward.
- Local retrieval: When using an on-device model ("search by meaning"), brain_ask and search_all_brains use BGE semantic ranking, supplemented by lexical help for exact identifiers, paths, and versions, running entirely locally. Without that model, retrieval is keyword-based only.
How to use
Requires Node.js 20 or higher. Run the install command inside the project; it creates brain.klypix (if the project does not have one), connects the locally detected editors, and ifit is a git repository, registers the .klypix merge driver, and exits only after a real MCP handshake has counted tool responses. After installation, reopen the coding tool or reconnect its MCP server, and allow the KLYPIX tools. You can run npx klypix-mcp doctor in the project to check the installation.
Typical flow: the agent calls brain_sync to start a session and declare the files it expects to modify, calls brain_note to record a decision or correction, and calls brain_ask to retrieve what the project currently believes. File overlap warnings are advisory and do not block edits.
When it applies
- Running multiple coding agents on the same codebase and needing shared project understanding.
- A team needs to version architecture decisions and their reasons alongside the code.
- Needing to track the supersession history of decisions and detect whether referenced code has changed.
Limitations
- In a standalone setup, brain data stays in your files, but the agent's provider may still receive what the agent reads.
- Standalone live presence and file overlap warnings are limited to the same machine and OS user.
- Integration depth varies by host.
- The semantic search model is about a 150 MB download and about 110 MB on disk; terminal-free installation requires an explicit --semantic to fetch it.
- The correction prompt itself does not guarantee a match; you need to confirm that the correction names an earlier decision.
- A small trial only checks the capture and recall paths; it does not establish the retrieval quality of a large brain, nor does it prove host automation.
