@jemidev/mcp

MCP server for Jemi. It lets an AI assistant - Claude Code, Codex, Cursor, and anything else that speaks the Model Context Protocol - read and edit your Jemi projects.

It forwards every request to a Jemi tab over a local WebSocket bridge on 127.0.0.1:4401.

assistant  ──stdio──▶  @jemidev/mcp  ──WebSocket──▶  your Jemi tab  ──▶  Jemi API
                    no keys, no storage               holds the keys

Usage

Register it with your assistant and let the assistant start it:

claude mcp add jemi -- bunx @jemidev/mcp    # or: npx -y @jemidev/mcp

Then open Jemi, go to Settings → MCP and press Connect. That page has ready-made snippets for Claude Code, Claude Desktop, Codex, OpenCode, Cursor, Gemini CLI, VS Code and Windsurf.

Several assistants can be registered at once. The browser can only reach one port, so the first process to start holds the bridge and the rest forward their calls to it — one tab serves all of them.

HTTP instead

The server also speaks Streamable HTTP on http://127.0.0.1:4401/mcp, for clients that cannot spawn a process. Then you start it yourself and leave it running:

bunx @jemidev/mcp
claude mcp add --transport http jemi http://127.0.0.1:4401/mcp

Which transport you get depends on whether stdin is a TTY: an MCP client spawns the process with a pipe and gets stdio, a human typing the command in a terminal gets the HTTP server. Pass --stdio or --http to say so explicitly.

Related MCP server: MCP Server Demo

Configuration

The origin allowlist matters: without it any page you visit could connect to the local port and feed your assistant fabricated project data. It could not read anything — it has no keys — but it could lie.

Tools

34 tools covering projects, channels, tasks, documents, comments, chat, dictionaries (tags, priorities, difficulties) and members. Ids are opaque, so resolve them first — jemi_list_projects → jemi_list_channels — before acting. The exception is a task number like ALP-0254: jemi_find_task_by_key resolves it in one call.

jemi_list_tasks filters by board, title, tag, priority, difficulty, assignee and by what a task is missing, and returns a brief shape by default — listing a busy channel in full does not fit in an assistant's context. Bulk tools (jemi_update_tasks, jemi_update_dictionary_items, jemi_create_tasks, jemi_delete_tasks) exist so twenty edits are one call.

Task bodies, comments and chat messages are markdown — CommonMark and GFM, plus a Jemi spelling for what markdown lacks (underline, colour, mentions, attachments). jemi_document_format returns the syntax each surface accepts.

Destructive tools (deleting a task, a channel, a dictionary entry) require a confirmation argument that the assistant must get from you in words. That is a guardrail against an over-eager agent, not a security boundary.

Bridge protocol

The frame types the server and the browser exchange are published as a subpath export, for anyone building another bridge:

import { MCP_PROTOCOL_VERSION, parseFrame, type TMcpBridgeFrame } from '@jemidev/mcp/protocol'

Only one bridge may be attached at a time: every operation targets "the tab the user is looking at", and with two tabs connected there is no way to say which one that is.

Development

bun install
bun run dev          # watch mode, from source
bun run build        # bundles dist/ for Node
bun run check-types

The runtime is plain Node (node:http + ws) so the package works under npx as well as bunx; Bun is used for the toolchain.

License

MIT © Jemi