rusty-harness

A tiny agent harness that's blazing fast. One Rust core runs the agent loop, serves it over A2A 1.0, and ships to Rust, Python and JavaScript:

• One agent, every surface. A terminal UI and a web renderer, both plain A2A clients of the same harness, with A2UI forms, images, voice and slash commands.
• Long-running tasks. A task stays open across turns, asks people for input through channels, plans its work as a DAG, and survives restarts on Temporal (or runs in process).
• Extensible. Tools, MCP servers, skills, plugins and hooks on every operation; models from OpenAI (Responses API) and Anthropic, or any OpenAI-compatible endpoint.
• Small and fast. About 260 µs of harness per turn, a 26 ms TUI cold start and a 20 MB binary.

Get	Install	Package
The rusty-harness command	npm i -g @rusty-harness/cli · pip install "rusty_harness[cli]"	npm · PyPI
Rust	cargo add rusty_harness --features rt-tokio	crates.io
Python	pip install rusty_harness	PyPI
JS / TS (WebAssembly)	npm i rusty-harness	npm

Quick start

export OPENAI_API_KEY=…   # referenced as env:OPENAI_API_KEY, never stored
rusty-harness             # writes ~/.rusty-harness/config.toml, starts the harness, opens the TUI
rusty-harness --no-tui    # the A2A server alone, on 127.0.0.1:7474

The demo runs the whole thing (the TUI and the web renderer on one harness, an A2UI ask, a task open across turns, Temporal Cloud) from a checked-in configuration.

Layout

All logic lives in crates/rusty-harness; the Python (crates/rusty-harness-py) and WebAssembly (crates/rusty-harness-wasm) crates are thin adapters, and web/ is the web renderer.

Documentation, including the architecture, the capabilities and local development for every layer, is at https://madarauchiha-314.github.io/rusty-harness/ and in docs/.