Desktop sandbox for experimenting with AI agent workflows locally.
Agent Sandbox Desktop is an incubating project intended to provide a safe local environment for running, observing, and debugging agent behavior before moving ideas into production tools.
The project is aimed at developers who want a dedicated workspace for:
- running agent tasks in isolated sessions
- inspecting tool calls, logs, and outputs
- testing prompts, policies, and agent loops
- comparing providers and configurations locally
- experimenting with multi-step automations in a desktop UI
- Task runner for agent jobs
- Session timeline and event log
- Tool execution console
- File and command sandboxing
- Provider and model switching
- Prompt/playbook presets
- Local-only storage for logs and runs
This repository is currently in early incubation. The public repo exists to hold the product direction, naming, and future implementation work.
Most agent tooling is either terminal-first or tightly coupled to one provider. Agent Sandbox Desktop is meant to be a local playground where agent behavior is observable, testable, and easier to iterate on.
- local-first agent experimentation
- safer tool execution workflows
- debugging and replay for agent runs
- desktop UX for agent builders
MIT