- Drawn by
- Lex
- Date
- 2025-11
- Scale
- 1 : 1
- Sheet
- 03 OF 05
- Tool actions pass through permission checks and, in governed mode, human approval before execution.
- Tools run in an isolated environment using gVisor or Docker, with controlled access to the workspace and network.
- The model’s context is assembled with source information, a size budget, and a content fingerprint for later inspection.
- Recorded runs and repeatable tests help check workflow behavior as the framework changes.
Scroll across the drawing to see the detail.
An AI agent combines a model with tools to work through a task. Relurpify is a system for building those agents. It provides the tools, execution rules, and context handling from which an AI agent workflow can be constructed. The central question is how to make that workflow inspectable and controllable by the person whose workspace it uses.
The framework is written in Go. It supports local models through Ollama, OpenAI-compatible model services, and a scripted backend for repeatable tests. Using a remote model service is a separate choice from running the framework locally.
Permission is part of execution§
Before a proposed tool action runs, the framework checks its permissions and expected effects. Governed mode adds a human approval step. Execution then takes place in a sandbox: an isolated environment with limits on workspace and network access.
These controls make the boundary between the model’s proposal and the actual action explicit. They also provide places to record and test what happened. The project’s documentation describes the available modes and security boundaries.
Context you can inspect§
The information supplied to the model is assembled into a recorded artifact. It includes where the material came from, a token budget that limits its size, and a digest, or content fingerprint. Record-and-replay support and test fixtures help investigate changes in behavior.
The framework includes Euclo, a coding agent; relurpish, its terminal interface; and tools for checking configuration and testing agent workflows. Recipes describe workflow behavior in a small purpose-built language.
What this brings together§
This work connects framework design, permissions, execution isolation, context preparation, and testing. It shows how common infrastructure can support different agent behaviors while keeping control points visible. The repository and project documentation carry the implementation detail.