Chidori

Chidori is an open-source AI agent framework for developers building long-running workflows that need to survive crashes or wait for human approval. It runs on your own machine or server, with binaries for macOS on Apple Silicon and Intel, and for Linux. The project uses the Apache 2.0 license.

Its main distinction is how it preserves a run. Chidori records model responses, tool results, and HTTP calls, then uses those records to reproduce the same output without calling the model again. Developers can revisit a failed run without paying for repeat model requests, or keep a checkpoint in Git as an integration test for changes in agent behavior. Exact replay requires the same code.

Runs save checkpoints to disk and can continue in a fresh process after a crash. An agent can also suspend while it waits for a person or another agent to respond, without keeping a process alive during the wait. Agents use ordinary async TypeScript rather than a graph editor or a separate workflow language.

The runtime is a single Rust binary with an embedded JavaScript engine, so it doesn't require Node, Deno, or V8. TypeScript and Python SDKs communicate with it over HTTP. It supports OpenAI-compatible backends, including Ollama, vLLM, LiteLLM, DeepSeek, and Groq. The runtime and checkpoints stay local; model requests go to the selected backend, which can run locally or use a cloud service. Built-in prompt caching marks stable prompt prefixes for provider caching.