ToolExperimental

jeju

Name: jeju
Rating: 3.0 (27 reviews)
Author: cosmtrek

by cosmtrek

Declarative local-first runtime for bounded, auditable AI agents

Updated Jul 31, 2026

Stars

Forks

View on GitHub

How It Works

Run bounded, local-first AI agents from a single manifest and record every side effect for audit. Jeju executes headless agent definitions (in Go) with strict boundaries on actions and resources, producing an auditable trace of effects and decisions. Its standout feature is declarative manifests that make agent behavior reproducible and inspectable across runs. Evaluation-Driven Development (EDDOps)

See related protocols

Standards this tool supports

The Value Proposition

As agents act autonomously, reproducible runtimes and immutable audit trails are essential to establish trust and diagnose failures. Jeju provides a lightweight local environment to constrain agents, capture their effects, and build an agent track record without cloud dependencies. Until now many agent experiments lacked a simple, declarative harness that combines runnable manifests with per-run effect logging for evaluation and governance. Guardrails Pattern

When to Use

Developers and researchers who need a reproducible, local harness to run and audit bounded agents during evaluation and pre-production testing. Hierarchical Multi-Agent Pattern This helps teams implement a structured approach to coordinating agents and maintaining a reliable agent track record using the Model Context Protocol (MCP) Pattern for tool integration.