What is Katari?

Katari is a programming language for orchestration: the layer of a system that decides which agent runs, with what input, what happens when one asks a human a question, and what survives a crash. You write that layer as a typed program; a persistent runtime executes it.

Three ideas carry the whole language:

• Agents are functions. An agent takes a labelled record in and returns a value out, and both sides carry a JSON schema. Anything can call it — another agent, an HTTP client, an AI model that discovered it as a tool.
• Effects are visible. An agent's signature tracks the requests it may perform (with ask). A request with a handler in scope is a function call; a request with no handler escalates — the run parks as an open question a human answers, minutes or days later.
• Execution is durable. The runtime persists every step. A run can wait on a webhook, a cron occurrence, or an OAuth authorization, survive a server restart, and resume exactly where it parked.

A taste

@"Ask a human to weigh in; the run waits until this question is answered."
request ask(question: string) -> string
 
@"Review one source by asking a human what stands out in it."
agent review(source: string) -> string with ask {
  let note = ask(question = f"What stands out in ${source}?")
  f"${source}: ${note}"
}
 
@"Review every source in parallel, then join the findings into one report."
agent main(sources: array[string]) -> string with ask {
  let notes = parallel for (let source in sources) {
    next review(source = source)
  }
  string.join(parts = notes, separator = "\n")
}

parallel for fans each source out to its own thread. Nothing handles ask, so every question escalates: the run parks, the console shows the blocked tree, and each answer — from the inbox or katari answer — resumes its thread. The compiler checked all of it: the effect row, the schemas, the join.

What ships with it

• A compiler that lowers Katari source to an IR, and a runtime server that executes IR snapshots against PostgreSQL — runs, escalations, and schedules all persist.
• A standard library (http, json, time, webhook, mcp, oauth, replay, …) and a package registry — ai (model-agnostic tool-calling over Anthropic / Gemini / OpenAI), discord, slack, google_calendar, tavily, and more. Every API is documented in the reference.
• MCP in both directions: consume any MCP server's tools as typed agents, or serve your agents as an MCP server.
• Tooling: a CLI, an LSP with a VSCode extension, and an admin web console with a live run trace.

Where to go next

• Installation — the CLI and a local runtime.
• Quickstart — a project running in five minutes.
• Tutorial — build a Discord bot that talks to a model and uses your agents as tools.
• Concepts — the execution model, one idea at a time.