Orchard logoOrchard

Roadmap

Orchard is a complete language with a working runtime, real model providers, and bindings for four platforms. The plan from here is to widen what agents can do, sharpen the tools around the language, and keep the whole thing fast and predictable. This page is honest about what exists today and what does not yet.

Shipped in 3.1

  • Token level streaming of the model’s reply, in the agent terminal and through the provider APIs.
  • The browser tool pack: drive a real headless Chrome to open a page, read it after JavaScript runs, click, type, and capture a screenshot.
  • The live agent terminal for orch run, with a slash menu of skills and tools and a streaming view of every model call and tool call.

Shipped in 3.0

  • The full language: parser, type system, static checker, and intermediate representation.
  • Typed generation with gen as T and exhaustive match.
  • The interpreter and runtime, including the autonomous delegate loop.
  • Concurrency with spawn, await, and parallel.
  • Tool packs for HTTP, the web, and the shell, plus a Model Context Protocol client.
  • Providers for Anthropic, OpenAI compatible endpoints, and Ollama, plus a mock provider for offline work.
  • Durable memory backed by redb, conversation recall, budgets, and a circuit breaker.
  • Triggers that run an agent on a schedule or when a file changes, and an orch serve mode.
  • The orch command line tool: check, compile, run, new, info, doctor, format, and serve.
  • Bindings for Rust, C, Python, and WebAssembly over one shared core.

In progress

  • A richer standard library of types and helpers that show up in everyday agents.
  • More detailed compiler diagnostics, with suggestions and clearer spans.
  • A growing set of golden examples and end to end tests across every binding.

Planned

  • A package format for sharing tool packs and reusable agents.
  • An editor experience with a language server, so checking and completion work as you type.
  • Tracing and replay, so a run can be recorded and stepped through afterwards.
  • First party adapters for more model providers and local runtimes.
  • A hosted playground that runs Orchard in the browser through WebAssembly.

Exploring

  • Static estimates of token and cost budgets, checked before a program runs.
  • Typed contracts between agents, so multi agent systems stay honest at their boundaries.
  • Deeper formal guarantees for the constrained generation path.

How we decide what comes next

Two questions guide the order of work. Does it make a correct agent easier to write, and does it keep the language small enough to hold in your head? Features that pull in both directions wait until we find a design that does not. Dates are deliberately absent here, because we would rather ship a feature when it is right than promise a quarter and rush it.

Shape the roadmap

The roadmap is public and the project is open source. If something you need is missing, the fastest way to move it up the list is to open an issue or a discussion on GitHub and tell us what you are trying to build. You can also read why Orchard exists to understand the principles behind these choices.