onPanda: open-source token-level LLM editor that can harness Claude Code and Codex
Fancy_Fanqi77 · reddit · 2026-09-19
After two years of development, the author open-sourced onPanda (on-Policy Alignment Data Annotator), an interactive tool for token-level LLM control, model inspection and data annotation.
Key features:
- Hover a token, click or freely edit alternatives, then continue generation — including reasoning and tool calls
- Edit prompts, branch tool calls, and record branch history in a tree structure for prompt engineering
- Multimodal support (images/video/audio), MCP connections for real-environment tasks
- Can harness Claude Code, Codex, and OpenCode to compare their tool sets, system prompts, skills, and memory
- Ships a browser-agent running in the browser with JS execution, retrieval, UI interaction, local file access, and persistent memory
Available online (mobile-friendly) and self-hostable via GitHub.
Related event: onPanda Open-Sourced: Token-Level LLM Editing Tool(2 posts)→
More from coding & agent
- ratex: pure-Rust TeX rebuild compiles docs 10-70x faster than TeX Live — deliprao · 2026-09-19
- Street Fighter poster animation on an 8GB GPU: MiniMax H3 + ComfyUI full breakdown — leftonredd33 · 2026-09-19
- AI coding agents' permission problem is becoming bigger than the model itself — Relevant-Potential17 · 2026-09-19
- Autonomous Labs open-sources OpenHarness, adds new senses to Lamp robot — dee_hw · 2026-09-19
- Six agent harnesses compared: token usage per task on identical model (Kimi K3) — zainhas · 2026-09-19
- Same model, 6 agent harnesses, 3x gap in end-to-end task completion time — zainhas · 2026-09-19