Open source · WebGPU · No backend

Inspect deeper. Understand more.

A keyboard-driven TUI for exploring LLM internals — logprobs, token distributions, and hidden-layer activations — running entirely in your browser via WebGPU. No server. No account. No setup.

Try the demoView on GitHub
logprobr — token inspect
❯inspect "The capital of France is"

top_token Paris91.4% Lyon4.3% Marseille2.1%…1.1%
logprob[0]-0.0898
entropy0.412 bits

❯layer 24 --head 07 --type attn

attn_pattern[0.00 0.01 0.03 0.82 0.09 0.04 0.01]
layer24 / 32 · head 07 / 32

❯

01 CAPABILITIES

Everything the model knows. Visible.

output_token_logprobs

Logprobs

Inspect per-token log-probabilities directly from the model's output distribution. See what the model almost said.

top_k_token_candidates

Token distributions

Visualize the full probability mass across token candidates at each generation step, not just the greedy pick.

residual_stream · mlp · attn_heads

Hidden-layer activations

Read residual stream, MLP, and attention head activations at any layer depth for any prompt.

attention_patterns · logit_lens

Mechanistic primitives

Attention pattern heatmaps and logit-lens projections to trace how information flows through the network.

02 HOW IT WORKS

Zero friction from URL to inference.

01

Open the TUI

Navigate to the site. No install, no CLI, no Docker. The interface loads entirely in your browser tab.

02

Load a model

Select a Hugging Face model. WebGPU pulls weights locally — they stay on your machine, never touch a server.

03

Inspect

Type a prompt. Use keyboard-driven commands to drill into logprobs, activations, and attention patterns in real time.

03 OPEN SOURCE

Fully open. Fork it.

The entire runtime is open source. No hidden inference, no telemetry, no account wall. Your prompts run locally on your GPU — we never see them.

Star on GitHub

Radical transparency

Expose what models actually do, not just what they output.

Zero-friction access

No server, no account, no setup. A URL is all you need.

Open by default

MIT licensed. Fork it, extend it, build on it.

Developer-first

Keyboard-driven. Built for people who read source code.