Steer LLMs and Agents at the Token Level: An interactive tool for token visualization & control, model inspection and data annotation.

Researchers introduce onPanda, an interactive tool for annotating LLM alignment data and agent trajectories via token-level correction. Annotators fix the first inappropriate token in a model response and let generation continue from the corrected prefix, producing data that stays on-policy while cutting annotation time. They also release Panda-CVL, an annotated dataset, plus a token-level correction benchmark.

OnPanda is an interactive web tool for visualizing tokens, editing model output at the token level, and annotating data. Users hover over a token, pick an alternative or edit it, and continue generation, including edits to reasoning and tool calls. It targets hackers, engineers, and power users who need deep model inspection and efficient data annotation.

Cartographer is an agentic knowledge base that generates and maintains Claude Code configuration — skills, subagents, hooks — and syncs it across a team. The LLM agent writes to the knowledge base under enforced conventions, avoiding broken links and lost history. It targets developers tired of hand-tidy KBs and of agent configs drifting across .claude/.codex/.opencode skill folders.

Products·building

TokenPrint: visual debugger for LLM inference

TokenPrint is an open-source visual debugger for transformer inference, similar to Chrome DevTools. It allows developers to step through a forward pass token-by-token, inspect activations, tensors, and KV cache, and visualize architecture components like GQA and RoPE. It solves the problem of debugging LLM inference interactively, which currently relies on logs and notebooks.

TokenPrint is an open-source visual debugger for transformer inference, similar to Chrome DevTools. It allows developers to step through a forward pass token-by-token, inspect activations, tensors, and KV cache, and visualize architecture components like GQA and RoPE. It solves the problem of debugging LLM inference interactively, which currently relies on logs and notebooks.

CORTEX is a Windows app that provides a visual interface for inspecting small local language models during generation. It shows live token probabilities, attention maps, hidden-state PCA, logit-lens views, system telemetry, and allows controlled interventions, making code-heavy interpretability workflows more accessible.