Introducing Choreographr

2026-08-01T00:00:00+00:00

Choreographr is an all-purpose, extensible AI agent system written entirely in Rust. It's early, it's open source (Apache-2.0), and it takes a few opinions that we think are worth sharing.</p>

One daemon, many clients</h2>
Most agents are a single process with a single UI. Choreographr is built the other way around: a daemon</strong> owns your sessions, and clients connect and disconnect whenever they want.</p>

`choreo-tui</code></strong> — a full-screen terminal UI with O(1) scrolling and streaming.</li>`
`choreo-gui</code></strong> — a desktop app (Dioxus), in development.</li>`
`choreo-im</code></strong> — chat with your agent from Telegram, in development.</li>`
choreo-acp</code></strong> — drive sessions from ACP-compatible editors like Claude Code.</li> </ul> Because sessions live in the daemon, you can start a task in the terminal and check on it from your phone. Everything persists in an embedded redb</code> database and survives restarts.</p> LLM-generated code runs in a sandbox</h2> Giving an agent direct shell access is one of the riskiest things you can do. Choreographr offers a different default: generated code is compiled to RISC-V and executed in an isolated VM with a 4 MB flat memory space and a configurable cycle budget. The guest can only interact with the world through registered tools — same registry, same credentials, same working directory.</p> That VM is also a tool-call scripting language: the agent can quickly write a little Rust snippet to orchestrate tools with custom logic.</p> 70+ providers, one catalog</h2> Choreographr speaks OpenAI-compatible, Anthropic Messages, and Google Gemini. Each provider is a data file in a TOML catalog — curated models, context windows, reasoning levels. Adding a new OpenAI-compatible provider is a data edit, not a code change.</p> High-performance by construction</h2> The daemon uses pure OS threads and message passing — no async runtime, no shared mutable state. Every thread owns its data. It's simple to reason about and fast.</p> What's next</h2> Subsessions (hierarchical subagents) already exist; git worktrees, cron, extensions, and OS-level sandboxing are on the roadmap. We'd love your help — star the repo, open issues, and join the Telegram community</a>.</p>