AgentDoggo Playground

AgentDoggo Playground is a desktop workspace for running and inspecting AI agents, designed to make experimentation easy and to show the machinery that is usually hidden behind a chat box. You can edit an agent’s prompt, context window, model settings, tools, and permissions, then watch the agent work in real time — seeing the exact request sent to the model, streamed output, tool calls and results, stopping conditions, and the state shared by the interface.

The Context View shows the exact payload sent to the model, including the system prompt, recent history, and enabled tools. The Graph shows the flow between the user, agent harness, model, environment, and store, with small JavaScript blocks that can be inserted at five points in that loop to rewrite or filter data. Permissions show folder grants and their 30-minute expiry directly in the interface, transcript replay lets you save a run and play it back later without a model or server running, and MCP connections let you point the Playground at any external Model Context Protocol server to inspect how a different tool environment changes the run.

The repository also ships two companion apps as teaching material: AgentDoggo World, a 9×9 gridworld exposed as an MCP server with a spectator view that gives agents a concrete environment to act in, and AgentDoggo Chat, a phone-style MCP chat server demonstrating that the Playground can connect to different kinds of servers. An offline test suite and optional live-LLM evaluations drive the same agent harness against the World server and check platform invariants such as tool reachability, state changes, and valid positions. Model calls are provider-agnostic, defaulting to OpenRouter with LM Studio and OpenAI-compatible endpoints also supported.

AgentDoggo Playground was developed as part of the lecture and workshop ‘Responsible Design for Agentic Systems’, held at CREATE Campus, Singapore as part of National Coding Week 2026.