SideButton: An Open-Source MCP Server for AI Coding Agents

Max Svistunov — Fri, 31 Jul 2026 15:02:11 +0000

SideButton is an open-source AI agent platform built around an MCP server. It gives any coding agent a real browser, a workflow engine, and reusable domain knowledge. Version 1.5 is out, and all of it runs on your own machine.

What is in the box

MCP server. Built-in tools for real-browser control, workflow execution, run logs, and artifacts. Claude Code, Cursor, Copilot, and Windsurf connect over stdio or HTTP + SSE, with MCP OAuth 2.1.

Workflow engine. Agentic workflows defined in YAML, with step types spanning browser actions, shell commands, LLM calls, and control flow, plus nested workflows and retries.

Knowledge packs. The domain knowledge that turns a general coding agent into a software engineer, QA engineer, or product manager for one specific app. Selectors, data models, workflows, and built-in roles per web app.

REST API and dashboard. A REST API so any webhook, cron job, or internal tool can trigger a workflow, plus a local dashboard with live run logs and a pack manager.

Tracker and chat steps work across Jira, Linear, GitHub, and Slack. The LLM provider is auto-detected from your environment.

Getting started

Install the server:

npx sidebutton@latest

Connect the SideButton Chrome extension, then point your agent at the server:

claude mcp add sidebutton --transport http http://localhost:9876/mcp

Install a knowledge pack for the web app you work on with sidebutton registry add, or scaffold your own with sidebutton init. Packs land in ~/.sidebutton/skills/ and load on startup.

From there the agent works on its own: running workflows, calling the LLM where reasoning is needed, logging every step. Your browser, your sessions, your API keys, all local.

Knowledge packs are just markdown

A pack is a folder of markdown files, so it reviews and diffs like code. You can ask Claude to capture what it learned during a session, or scaffold a pack with sidebutton init, then share it through a registry that is either a local folder or a git repo. Installed packs merge their workflows, roles, and domain targets into your local config at startup.

Plugins add your own tools

The MCP server is extensible in any language. A plugin is a folder with a manifest and a handler script under ~/.sidebutton/plugins/, loaded alongside the built-in tools at startup. Two ship in the box: screen-record, which captures an agent session to MP4, and writing-quality, which scores content before it goes out.

What is next

Native iOS and Android apps, and a public registry so teams can discover and share knowledge packs instead of rebuilding them.

The platform is stable and open source today. Browse the AI agent skills already published for popular web apps, or write one for the app your team lives in.