A Journal-First Runtime That Makes Agents Show Their Work

What it does

Kungfu is a low-latency, append-only event journal with a shared, strongly-typed schema, designed to capture, share, and faithfully replay high-frequency event streams. It exposes journal data zero-copy to C++, Python, and Node, and wraps the core in SDKs and a minimal reference application. Originally created for trading execution, it now targets agent-facing applications that need deterministic replay and local accountability.

The interesting bit

The project is built on the principle that facts come before trust: every load-bearing claim should be bound to a verifiable source, artifact, or runtime receipt, and the same runtime handles both live execution and deterministic replay from identical journal semantics. It also deliberately absorbs toolchain complexity by embedding Python and Node, so most extension development needs no separately installed interpreter or package manager.

Key highlights

• Journal-first data plane (yijinjing) with a unified type system (longfist) carrying nanosecond timestamps, source, destination, and message type.
• Zero-copy, multi-language in-process sharing across C++, Python, and Node without hot-path serialization.
• Deterministic replay using the same runtime and journal semantics for live and recorded streams.
• Batteries-included toolchain: the runtime embeds Python and Node, aiming to remove external dependency management for most extension development.
• Cross-platform: runs on macOS, Linux, and Windows (including arm64).

Caveats

• The repository is explicitly labeled “coming soon” and currently contains only the open-source runtime core, SDK, and reference applications being prepared in public.
• The full commercial product surface has not yet shipped.

Verdict

Developers building agent tools or high-frequency event pipelines that need deterministic replay and local auditability should look here; those wanting a finished, turnkey agent product should wait for the commercial surface to arrive.