Hi everyone! 👋

I started building onPanda in 2024, and over the past two years it has gradually grown into a workspace for deeply exploring, steering, and annotating LLMs and agents.

The core interaction is simple: when you want to change where a generation is heading, select a token, inspect its alternatives or edit it freely, and continue generation from that point. This makes it easy to see how a small intervention can change everything that follows.

I wanted this interaction to work beyond normal chat output. In onPanda, you can do token-level correction directly inside the generated text, including reasoning and tool calls, while keeping branches and their history. You can also connect tools, MCP servers, and agent harnesses such as Claude Code, Codex, and OpenCode to inspect and modify more complex agent trajectories.

The name "onPanda" originally comes from "on-Policy Alignment Data Annotator." Data annotation is still an important part of the project: token-level corrections can produce precise supervision while keeping the corrected trajectory close to the model’s own policy. In a small controlled study, this workflow reduced median annotation time by 52% compared with manual post-editing.

But onPanda has grown beyond annotation. I now use it as a general interface for looking inside model generation, experimenting with alternatives, debugging agent behavior, and understanding what changes when you intervene at a particular point.

The project is fully open source and self-hostable. I’d love to hear how you would use this kind of interface—especially what you’d want to inspect, compare, or control when working with LLMs and agents.

Thanks for checking it out 🐼

Try it online (mobile-friendly): https://onpanda.diyer22.com

onPanda

Inspect and steer LLMs and agents at the token level

Inspect and steer LLMs and agents at the token level