What it does
coldstart keeps two layers of codebase knowledge for AI coding agents. The first is a notebook of durable notes about how a repository works: what a file is for, how a flow spans several files, and which invariants hold. Notes are captured after real tasks, recalled when a later task matches, and each one is anchored to real files, so a note whose evidence has drifted is flagged rather than served as current. The second layer is a static index over file paths, symbol names, exports and the import and call graph. It is used to find which files are relevant to a task, and it returns checkable evidence rather than a similarity score.
It is relevant to developers who use Claude Code, Codex or Cursor and want agents to stop re-reading the same repository at the start of every session and re-deriving what an earlier session already established. After an agent edits two different files, a PostToolUse hook also suggests files that git history shows changing alongside them, such as a sibling implementation, a test or a doc, and asks the agent to check whether the link is real.
Running it requires Node.js 18 or later. A single init command sets up navigation and the notebook for the client the user selects, writing guidance into that client's rules file and registering hooks in its settings. The notebook is stored inside the repository under .coldstart/notebook/. An unwire command removes only coldstart-owned entries from each project and keeps the notebook by default.
Limitations and requirements
Requires Node.js 18 or later. The client must be chosen explicitly during init; it is not auto-detected. The mcp experience additionally writes MCP configuration files into the project. Hooks and rules are written into project files such as .claude/settings.json, AGENTS.md and .cursor/hooks.json, and unwire must be run in each project before the global npm package is uninstalled.
Safety
Our team has not reviewed this resource. Only automated rules checked it; resources with high-risk signals are not published. These are signals, not a guarantee that installing it is safe. Try it in a sandbox project first.
- Runs code or changes an assistant's behaviour
Sources
- Official MCP Registry first seen 13 Nov 2026 · last seen 13 Nov 2026 · its page on the source
coldstart
Codebase memory for AI agents: an AST index plus agent-written notes that self-stale.
v2.2.17As shown on registry.modelcontextprotocol.io when we last read it