Show HN: OnPanda – Steer LLMs and agents at the token level
onpanda.diyer22.com
onPanda is designed for hackers, power users, curious minds, and engineers. Its UI is built for deep exploration and efficient data annotation.
- The core loop is simple: hover over a token → click an alternative or edit freely → continue generation. You can edit every part of model output exposed by onPanda, including reasoning and tool calls.
- Edit prompts directly, branch tool calls, and use a tree structure to record branch history. This makes onPanda useful for model inspection and prompt engineering.
- Support multiple modalities, including images, video, and audio; use tool calls and connect MCP servers to perform tasks in real environments.
- Connect popular harnesses such as Claude Code, Codex, and OpenCode to execute tasks. Explore and compare their tool sets, system prompts, skills, and memory mechanisms.
- onPanda includes a built-in browser-agent, an agent that runs in the user's browser without installation. It uses the browser as its harness and provides JavaScript execution, information retrieval, interface interaction, multimedia I/O, local file access, and persistent memory.
- onPanda stands for on-Policy Alignment Data Annotator.
- As an annotation tool, onPanda efficiently labels on-policy data. Its token-level correction format also provides fine-grained supervision with precise positions and naturally paired positive–negative samples.
- We believe these unique properties will make token-level correction a highly efficient and practical paradigm for future LLM alignment.
- Our paper and benchmark for token-level correction: https://on-panda.github.io/research/
Try it online (works on mobile): https://onpanda.diyer22.com/
GitHub repo for self-hosting: https://github.com/on-panda/on-panda
Detailed introduction on X: https://x.com/diyerxx/status/2101020850405462041