AgentGUI

AgentGUI gives researchers and developers a visual workspace for supervising long-running AI agents. Its pixel-art office represents tasks as desks, so you can follow several jobs and intervene while they run.

• Live task inspection: Review reasoning, tool calls, generated files, conversations, and terminal output, with a breakdown of time spent reasoning, using tools, or idle.
• Auditing and steering: An autonomous manager checks work against an editable task description and requests another pass when checks fail; you can also interrupt, redirect, or reassign agents.
• Hermes teams: Import existing Hermes agents, customize their personas and memory, organize tasks by topic, and share files across the team.
• Sandboxed actions: Agents run terminal actions inside Docker sandboxes, with tool access limited to what each assignment needs.
• Saved runs: Load earlier trajectories to review an agent's actions and results after a session ends.

The app and inference can run entirely on your own hardware, using Ollama on a laptop or vLLM on a local server. Optional Gemini and Claude API backends require API keys and send inference requests to those providers. Tasks that use web tools or external APIs still depend on those services. AgentGUI is open source under the MIT license and requires Hermes and Docker.