Chidori — Deterministic Agents as Code

Chidori is a YAML-free AI agent framework where agents are plain TypeScript — native async, typed inputs, and ordinary imports, no DSL — with checkpointing, replay, and event-driven execution built in.

Every side effect — LLM calls, tool invocations, HTTP requests, human input — goes through a fixed set of host functions on an injected chidori object the runtime can log, cache, and replay. Save a session's call log to disk, replay it later for identical output with zero LLM calls.

Why Chidori?

• Agents are TypeScript. Native async, typed inputs, ordinary imports — no template DSL, no YAML.
• Deterministic execution. Durable runs use a fixed Date, seeded Math.random, and route every side effect through a logged, cacheable host function call.
• Zero-cost checkpointing. Save a session's call log, replay it for the same output at zero token cost.
• Event-driven. Run an agent as an HTTP server that reacts to webhooks, alerts, or custom events.
• Rust core, TS + Python SDKs. The runtime is a single binary. The SDKs are pure HTTP clients to chidori serve — the Python SDK is pure-stdlib, with no dependencies.

A first agent

// agents/summarizer.ts
import type { Chidori } from "chidori";

export async function agent(input: { document: string }, chidori: Chidori) {
  const summary = await chidori.prompt("Summarize in 3 bullets:\n" + input.document, { type: "summary" });
  const actionItems = await chidori.prompt("Extract action items:\n" + summary, { type: "actions" });
  return { summary, actionItems };
}

chidori run agents/summarizer.ts --input document=@notes.txt