an autonomous AI coding agent orchestrator

what are beads?

a bead is a unit of work — small enough for an AI agent to complete in one session. think of it as a task, issue, or ticket: "fix the auth bug." "add retry logic." "write the migration."

beads live in your repo as a dolt database (version-controlled SQL). your data stays in your repo — not in a SaaS platform you don't control.

also known as: task, issue, work item, ticket

how does rosary automate AI coding agents?

rosary runs a reconciliation loop (like a kubernetes controller for code tasks):

1. scan — find open beads across your registered repos
2. triage — score and prioritize (dependency readiness, age, retry history)
3. dispatch — assign an agent to work in an isolated git worktree
4. verify — compile, test, lint, diff sanity check
5. close — open a PR or deadletter for human attention

agents use mache (structural code intelligence) to understand the codebase before making changes, and signet to sign their commits so you know which agent made which change.

how does rosary compare to Devin, Cursor, or GitHub Copilot Workspace?

Devin is a hosted AI software engineer — one agent, one task, cloud-only. rosary dispatches multiple agents across multiple repos simultaneously, and you can self-host it.

Cursor Background Agents run in Cursor's cloud. rosary is editor-agnostic — it works with Claude Code, Cursor, or any MCP client. bring your own editor.

GitHub Copilot Workspace is tied to GitHub's ecosystem. rosary works with any git host and any AI model (Claude, Gemini, or your own). bring your own everything.

the key difference: rosary is open source, runs on the MCP protocol, and gives you full control over your agents, compute, storage, and identity.

what is MCP and why does rosary use it?

the Model Context Protocol (MCP) is an open standard for connecting AI models to tools and data. rosary exposes 27 MCP tools over streamable HTTP — bead CRUD, agent dispatch, pipeline queries, workspace management. see the endpoint →

because rosary speaks MCP, any MCP-compatible client can use it: Claude Code, Cursor, Windsurf, or your own tools. no vendor lock-in, no proprietary API.

is rosary open source?

yes. rosary (orchestrator, rust), mache (code intelligence, go), and signet (identity, go) are all open source.

you can self-host the entire stack. rosary.bot is the managed version — same tools, hosted on the edge. pay if you want it managed. run it yourself if you don't.

how do I get started?

hosted (one command):

claude mcp add --transport http rosary https://rosary.bot

opens browser for OAuth. no install needed.

local (self-hosted):

brew install rosary && rsry serve
claude mcp add --transport http rosary http://localhost:8080/mcp

runs on your machine. your data stays local. unlimited.

what tools make up rosary?

rosary is three open-source tools that compose together:

what can I use rosary with?

any AI model: Claude, Gemini, or custom agents via the Agent Client Protocol.

any editor: Claude Code, Cursor, Windsurf, or any MCP client.

any compute: local subprocess, sprites.dev, Cloudflare Containers, or Fly Machines.

any git host: GitHub, GitLab, Bitbucket, self-hosted.