onPanda: An Interactive Tool for Token-Level LLM Steering and Agent Inspection
Developer Fancy_Fanqi77 has launched onPanda, an open-source tool designed for token-level inspection, output editing, prompt branching, and steering of LLMs and AI agents. The tool enables users to modify tokens mid-generation, trace reasoning and tool calls, and interact with models via an interactive tree structure. Debugging complex AI agents and fine-tuning prompt behaviors often requires granular visibility into how a model makes step-by-step decisions. By providing token-level control, multimodal capabilities, and harness integrations, onPanda simplifies model inspection, data annotation, and agent alignment for developers. onPanda features a built-in browser agent that runs locally without installation to execute JavaScript, manage files, and maintain persistent memory. It also connects with Model Context Protocol (MCP) servers and external agent harnesses like Claude Code, Codex, and OpenCode to compare tool sets and memory mechanisms.
## BACKGROUND
Large language models generate text by predicting one token—a sub-word chunk of text—at a time. An agent harness provides the runtime structure, system prompts, and execution loops needed for an LLM to take actions and interact with external systems. Additionally, the Model Context Protocol (MCP) is an open standard designed to standardize how AI applications integrate with external tools and data sources.