AI agents forget everything
You spent an hour explaining your architecture. The agent wrote perfect code. Next morning, new chat, blank slate. You start over. Again.
^ goldfish with PhDs
One MCP server for files, images, PDFs, web, and workflows. Your files stay local. No telemetry. No account. No cloud. Single binary. Zero config.
^ open source, free forever
PM passes to Dev, Dev passes to QA, QA ships it ✨
You spent an hour explaining your architecture. The agent wrote perfect code. Next morning, new chat, blank slate. You start over. Again.
^ goldfish with PhDs
Relay gives your agents persistent memory and structured handoffs. Agent A checkpoints its work, Agent B picks up exactly where A left off. Like a relay race, but for shipping software.
context flows, never lost
Each one has a role. Together, they ship products.
Breaks down features, creates issues, writes specs
Writes code, runs tests, opens pull requests
Reviews diffs, catches bugs, approves merges
Deploys, writes changelogs, announces launches
One command. Mac, Linux, Windows. Zero dependencies.
npx userelay
Works with Claude Desktop, Cursor, Windsurf, VS Code, or any MCP tool.
npx userelay init
40 tools across 7 categories. Files, images, PDFs, web, and more.
npx userelay status
$ userelay status relay ✓ v0.3.0 running tools ✓ 40 registered (7 categories) transport ◌ stdio (default) ✓ files, images, pdfs, web, data your files stay local. no cloud.
Your agents keep working after you log off. Wake up to merged PRs, updated docs, and a shipped product. Seriously.
MCP (Model Context Protocol) is a standard that lets AI tools like Claude Desktop and Cursor connect to external servers for extra capabilities. Relay is one of those servers. You plug it in, and your AI gets access to tools, memory, and coordination it didn't have before.
Nope. You install the binary and connect it. It works with whatever AI client you already use. If you want to contribute or extend it, then yes, the codebase is Go.
AI models forget everything between chats. Relay gives them persistent memory (checkpoints), the ability to hand off work to other agents (orchestration), and real-world capabilities like resizing images or parsing PDFs (tools). Think of it as upgrading your AI from a chatbot to a coworker.
MIT licensed, fully open source, no telemetry, no paid tier planned. We build in the open because we use it ourselves and want others to benefit. The "catch" is it's early and we'd love your contributions.
Yes. Any MCP-compatible client works. Claude Desktop, Cursor, Windsurf, and even terminal-based clients. If it speaks MCP, it connects to Relay.
Yes. Any MCP-compatible client works. Claude Desktop, Cursor, Windsurf, VS Code, and even terminal-based clients. If it speaks MCP, it connects to Relay. Run npx userelay init to set up your editor automatically.
Check out our contributing guide. We have open issues tagged for first-timers. The codebase is clean Go with zero magic. Pick a tool that interests you and send a PR.
Relay runs 100% locally on your machine. No cloud, no analytics, no telemetry. Your checkpoints and agent data stay in your project directory. We never see them.
Open source. Written in Go. Free forever.
Get Relay on GitHub ^ takes 30 seconds, promise