Mohak Bajaj has released an MCP (Model Context Protocol) server for his Mac note-taking app, Jot, allowing AI coding assistants like Claude Code, Codex, and Cursor to directly read, search, and append to a local Markdown vault. This integration solves a persistent friction point for developers: the tedious manual copy-pasting of context from external notes into agent prompts. The server runs locally over stdio, ensuring that notes remain on the userβs machine unless explicitly sent to the model by the agent.
Seamless Integration Across Major Agents
The setup is designed for minimal friction, requiring only a single command to link the CLI and add the server to the agent's configuration. For Claude Code, users run claude mcp add --scope user jot -- /usr/local/bin/jot mcp, which makes the tool available across all projects. Codex users can similarly add the server with codex mcp add jot -- /usr/local/bin/jot mcp, while Cursor and other JSON-configurable clients can paste a standard MCP server configuration block. This uniformity highlights the growing maturity of the MCP ecosystem, where tool compatibility is becoming a standard feature rather than a custom hack.
Seven Tools for Context Management
The Jot MCP server exposes seven distinct tools: list_jots, search_jots, read_jot, add_jot, set_done, archive_jot, and tag_jot. These tools allow agents to filter notes by tags or status, perform case-insensitive searches on body text and metadata, and read entire Markdown files including front matter and image paths. Crucially, the write operations (add_jot, set_done, etc.) use atomic writes to the underlying Markdown files, ensuring that the Jot appβwhich watches the folderβimmediately reflects any changes made by the AI agent.
The Safety of No Delete
A deliberate design choice in the Jot MCP server is the absence of a delete_jot tool. Bajaj argues that agents act quickly and can make mistakes, so destructive actions should be reserved for humans. Instead, the server offers archive_jot, which moves a note out of the active inbox but retains the file in the directory. This approach prioritizes data preservation and reversibility, preventing scenarios where a vague prompt like "clean up my old notes" results in permanent data loss. Users can still delete notes manually via the app, which sends files to a trash folder.
Key Takeaways
- Jot's MCP server enables Claude Code, Codex, and Cursor to read and write local Markdown notes without manual pasting.
- The server runs locally on macOS via stdio, preserving data privacy by keeping notes on the device unless sent to the model.
- There is no delete tool; agents can only archive notes, preventing accidental data loss from ambiguous prompts.
- Setup requires linking the Jot CLI and adding the server to the agent's MCP configuration with a single command.
- Agents can search, list, read, add, tag, and mark notes as done, creating a bidirectional flow between human notes and AI context.
The Bottom Line
Jotβs MCP integration is a pragmatic step toward reducing context-switching friction for developers. By enforcing a 'no-delete' policy, it wisely acknowledges the current limitations of autonomous agents, making it a safe and useful addition to the local development stack.