The agent language · declarative & standalone

Write agent and device routines in Kaze.

Kaze is a domain-specific scripting language for bounded agent and device workflows. The parser, VM and compiler are available. The current embedded host can produce synthetic AI, tool and memory results; a real model-and-authorized-tool journey remains an open milestone.

Domain-specific language · Capability declarations · Standalone compilation · Experimental integrations.
# Pure arithmetic example — no external capabilities
action add(a, b):
return a + b
How it works

From declarative script to native executable.

Check and compile supported routines. Real capability integrations require further implementation and testing.

1

Write Declaratively

Express agent routines, input parameters, and capability gates (needs: ai, tool, policy) in pure agent.kaze. Zero auxiliary wrapper scripts.

2

Compile to Single Binary

Run sofuu compile agent.kaze -o my-agent to package a supported routine. OS libraries and workload-specific resources may still be required.

3

12-Factor Configuration

Pass parameters through CLI flags (--arg), JSON config (--config), or environment variables (AGENT_*). Never hardcoded.

4

Supervisor Mode

The supervisor supports repeated execution and signal handling. Uptime and recovery depend on the workload and environment.

# Check a Kaze routine without executing capability calls
$ sofuu script check agent.kaze
# Package a supported routine as a native executable
$ sofuu compile agent.kaze -o my-agent
FAQ

Frequently asked questions.

Key technical facts about the Kaze language, standalone compilation, and security model.

What is Kaze?
Kaze is a domain-specific language for bounded agent and device routines, not a general-purpose language.
Does Kaze require Node.js or Python?
Supported routines can run through Sofuu or be compiled without a separate Node.js or Python installation. Operating-system libraries and workload-specific services may be needed.
Can Kaze call real models and tools today?
The end-to-end real model-and-authorized-tool journey remains an open milestone. The current embedded host can return synthetic AI, tool, and memory results; do not treat those results as real actions.
Can a supervisor guarantee uptime?
The supervisor provides repeated execution and signal handling. No universal uptime or crash-free guarantee has been established.
How do capability declarations help?
Kaze supports declared capability families and checks. Their effectiveness depends on the host and implementation; they are not proof that every effect is protected or every secret detected.