SideButton

SideButton MCP Server — Model Context Protocol for AI Coding Agents

SideButton is an open-source Model Context Protocol (MCP) server. It exposes agent tools — browser automation, a YAML workflow engine, and installable knowledge packs — to any MCP-compatible coding agent. One server, every client.

How the SideButton MCP Server Works with AI Coding Agents

An MCP topology has three layers. Your MCP client — an AI coding agent like Claude Code, OpenAI Codex, or Cursor — speaks the protocol over stdio or SSE. The SideButton MCP server runs on your machine, translates tool calls into concrete actions, and talks to the Chrome extension for browser control, the local filesystem for workflow execution, and the knowledge-pack registry for app-specific context. No traffic leaves your machine unless a pack explicitly reaches out.

Because SideButton runs entirely on your machine, you own the whole pipeline end to end. Audit every tool call, inspect every prompt, and extend packs without waiting on upstream reviews — the kind of ownership most AI agent platforms do not offer.

Supported MCP Clients — Claude Code, OpenAI Codex, Cursor, Claude Desktop, ChatGPT

Pick your coding agent for a setup walkthrough. All clients work — some just don't have a dedicated guide yet.

Setup guide

Claude Code MCP Server

Install SideButton in Claude Code. Browser automation, knowledge packs, and workflow engine available as MCP tools inside every Claude Code session.

Setup guide Works today with stdio transport.

OpenAI Codex MCP Server

Add SideButton to the OpenAI Codex CLI via MCP. Browser and workflow tools, TOML config, skill packs for any web app.

Generic MCP

Cursor MCP Server

Cursor reads MCP servers from ~/.cursor/mcp.json. SideButton works today with the generic MCP config — a dedicated walkthrough is on the way.

Generic MCP

Claude Desktop MCP Server

Claude Desktop picks up MCP servers from its developer settings. SideButton attaches via stdio — dedicated walkthrough coming.

Generic MCP

ChatGPT MCP Server

ChatGPT recently added MCP support for custom connectors. SideButton exposes its tool surface via SSE — setup guide coming.

Quick Start — Install the Open-Source MCP Server in 60 Seconds

Install globally, then run sidebutton. The bare command boots the server with the dashboard and every transport enabled.

$ npm install -g sidebutton
$ sidebutton

You should see:

  SideButton

  Config: ~/.sidebutton

SideButton server started on http://localhost:9876
  - Dashboard:  http://localhost:9876
  - WebSocket:  ws://localhost:9876/ws
  - MCP:        http://localhost:9876/mcp
  - API:        http://localhost:9876/api

Press Ctrl+C to stop

What each URL is for

http://localhost:9876/mcp

The MCP endpoint. Paste this URL into your coding agent's MCP config — Claude Code, OpenAI Codex, Cursor, Claude Desktop, or ChatGPT. Supports HTTP Streamable and SSE transports; the client picks automatically.

http://localhost:9876

The Dashboard. Open in a browser to inspect installed skill packs, browse available MCP tools, edit workflows, and watch live session logs.

ws://localhost:9876/ws

The WebSocket bus. Powers the live dashboard view and the bidirectional link to the SideButton Chrome extension that drives real browser sessions.

http://localhost:9876/api

The REST API. Programmatic access for automations, CI jobs, custom dashboards, or anything that needs to call SideButton without going through an MCP client.

Next: point your coding agent at the MCP URL. See Claude Code or OpenAI Codex for client-specific config.

What SideButton Ships — AI Agent Tools, Knowledge Packs, Workflow Engine

AI Agent Tools & MCP Tools

Browser control, element interaction, navigation, network observation, filesystem, and shell primitives — every one callable from an MCP prompt.

Installable knowledge packs

One command installs a whole app's selectors, data model, and workflows. Packs are public domain knowledge — teach agents GitHub, Jira, Linear, LinkedIn, or any site.

Agentic Workflow Engine

YAML-authored pipelines with 40 step types — conditionals, loops, LLM calls, browser ops. Agents invoke workflows like functions.

Chrome Extension & Browser Automation

A local Chrome extension exposes the real browser to MCP calls. No headless quirks — agents see the same DOM your Cursor does.

Self-Hosted AI Agent Platform

Runs entirely on your machine. No cloud dependency, no account, no phone-home. Your prompts and data stay local.

Open-Source MCP Server (Apache 2.0)

Audit every line, fork freely. Published to npm as sidebutton, source on GitHub.

What is the Model Context Protocol?

The Model Context Protocol is an open JSON-RPC specification introduced by Anthropic in late 2024. It standardizes how LLM clients discover available tools, call them with typed arguments, and stream results back into the conversation. Before MCP, every coding agent shipped its own tool-calling format — Claude Code had one, OpenAI Codex had another, Cursor had a third. MCP collapses that into a single protocol.

Practically, MCP means a coding agent on your laptop can reach any compliant server — local, remote, or bundled — without integration code on the client side. That changes the economics of tool building. Ship one MCP server, reach every agent. SideButton is that server for browser-driven agentic work: the kind of coding that spans editors, issue trackers, deployment dashboards, and the open web.

Frequently asked questions

What is the Model Context Protocol?

The Model Context Protocol (MCP) is an open JSON-RPC spec introduced by Anthropic in late 2024. It defines how large language model clients discover and call external tools, fetch resources, and receive streamed results. Any MCP client can connect to any MCP server, so one server can power Claude Code, Cursor, Claude Desktop, ChatGPT, and OpenAI Codex at the same time.

Which MCP clients does SideButton work with?

SideButton supports every MCP 0.1+ client over stdio or Server-Sent Events. Verified today: Claude Code, OpenAI Codex CLI, Cursor, Claude Desktop, and ChatGPT custom connectors. Any future client that speaks the spec will work without SideButton-side changes.

Is SideButton free and open source?

Yes. SideButton is Apache 2.0 licensed, published to npm as the sidebutton package, and developed in public on GitHub. Self-host on any machine with Node.js 20 or later. No telemetry, no license keys, no account required.

How do I install a knowledge pack?

Run sidebutton install <domain> — for example, sidebutton install github.com. The pack is fetched from the public registry, validated, and loaded on next server start. Browse the full catalog at /skills or author a new one at /create-agentic-ai-skill.

How is SideButton different from writing a custom MCP server from scratch?

A custom MCP server means implementing every tool yourself, shipping your own browser layer, and maintaining transport code. SideButton arrives with a full set of MCP tools, a workflow engine, Chrome-extension browser control, and a knowledge-pack marketplace. You write packs instead of servers.

Related