Chidori documentation
Docs track the next release. These pages describe the code on the default branch, which may include features not yet in the release you installed. When running a tagged release, read the docs at that tag (
https://github.com/ThousandBirdsInc/chidori/tree/vX.Y.Z/docs) and trustchidori <command> --helpover any page. Features that shipped after the latest release are marked with a "since" note where the skew bites.
Chidori is the agent framework where every run is durable, replayable, and
resumable by default. You write agents as plain async TypeScript — real
if/for/try, real imports, no graph DSL — and run them on a single Rust
binary with no Node, no Deno, and no V8. (The TypeScript and Python SDKs
are optional HTTP clients for driving a served runtime from your app —
agents themselves are always TypeScript. More FAQ →)
The one mechanism
Everything Chidori does follows from a single boundary: an agent never touches the world directly. Every side effect — every LLM call, tool call, HTTP request, file write, human question — flows through the runtime as a recorded host call. Because the runtime sees and records everything, it can log it, cache it, replay it, pause on it, and resume from it:
| Because every effect is recorded… | …you get |
|---|---|
| The journal is a deterministic record | Replay any run with zero LLM calls — byte-identical output, no tokens billed (Replay & Resume) |
| Runs persist at every host safepoint | Crash recovery — kill the process mid-run, resume in a new one (Durable Storage) |
| A pause is just a host call with no answer yet | Humans in the loop without a live process — suspend to disk, resume days later (Signals) |
| A recording fully specifies behavior | Recorded runs as CI tests — chidori verify asserts zero drift for $0 (Replay & Resume) |
| Code changes are events too | Implementation history — a git-like chain of every source version that ran, diffable, anchored to the journal (Source History) |
Where to start
- Getting Started — install the binary, scaffold and run your first agent.
- Your First Agent — a fifteen-minute tutorial: write an agent, pause it for approval, replay it for $0, check it into CI.
- Core Concepts — the host-function surface and the mental model.
- Common Patterns — which primitive fits which job.
Then pick up the rest of Using Chidori as you need it, and keep the Host API Reference and CLI Reference at hand. Evaluating against other frameworks? Start with the FAQ and the comparison table.
How these docs are organized
- Using Chidori — guides for agent authors and operators, roughly in reading order.
- Reference — the complete host API and CLI, for lookup.
- Internals — how the engine and runtime work, for the curious and for contributors.
- Posts — longer-form writing about the ideas behind the framework.
Deeper engineering notes and the historical usability reviews live in the
repository under docs/ rather than on this site.
Other references
llm.txt— the complete API reference, optimized for LLMs generating agents.- TypeScript SDK and Python SDK — HTTP clients with no native bindings.
- Examples — runnable agents, from hello-world to multi-agent war rooms.
- Discord — questions, ideas, contributions.