CyCast

CyCast for macOS

Beauty of AI.

Run open-weight models on your Mac and see the beauty in how they work: each token, layer, attention pattern, and next-token choice as an answer takes shape.

Your prompts, replies, models and recordings never leave your Mac. The app sends no analytics and no telemetry. Downloading a model contacts Hugging Face; see Privacy.

To sign in, CyCast keeps your email, name, a hashed password and the Macs you use it on. Payments are handled by Stripe. What CyCast keeps →

Download for MacSee pricing → Apple silicon · macOS 26.2 or later · CA$20 per year for 2 Macs · updates included while subscribed

Inside CyCast

A window into the model.

An illustration of what CyCast shows on your Mac: the input tokens, a path through the transformer layers, and the candidates for the next token.

Interactive illustration · no data
CyCastIllustration · no model runs in your browser
CyCast
DashboardAI Processing Behavior AnalysisCompare
gemma-4-E4B-it-MLX-4bit

Observe the model’s internal operations

gemma-4-E4B-it-MLX-4bit · deep capture

Overview Layer by layer
Input tokens
·primary·color.<turn|>↵
Model pathDrawn by hand
first layerlast layer
Residual stream AttentionMLP
Next token candidates
<turn|>
.
·(
▶ ReplayToken probabilitiesAttention map

An illustration of the CyCast interface, not a screenshot: it shows no data from a model. In the app, a run's data comes from a compatible model running locally on your Mac.

Every number says where it came from

measured
Read directly from the model while it runs.
derived
Calculated from measured values.
interpreted
A reading of the data, marked as one.

Deeper capture, same answer

Standard and Deep capture leave the output exactly as Light capture writes it, bit for bit, in tests with greedy decoding on SmolLM-135M and Gemma 4 E4B (MLX) and SmolLM-135M (PyTorch).

Your work stays on your Mac

Models, prompts, replies, and recordings never leave the machine. Your account holds your sign-in, your purchase, and the Macs you use CyCast on.

Future direction · Planned

Wireshark for AI

Capture, inspect and diagnose AI execution.

CyCast is growing toward an execution analyzer for local AI. Our goal is to capture what actually happens inside a running model, then let you inspect one inference as closely as a network capture. This is the direction for a future release of CyCast Trace.

AI should leave evidence of its execution.

As AI systems gain tools and autonomy, we want to connect model activity with the actions that follow it. A recording should help you reconstruct a run, compare it with another and see where it diverged. Read the CyCast vision →

Capture

The run that matters

Record the model and runtime you actually use, including quantized GGUF models. The first external target is llama.cpp; Ollama is a later target, alongside CyCast’s MLX and PyTorch engines.

Inspect

Follow the execution

Move through an event timeline. Filter by token, layer or latency, then inspect the selected event through attention maps, tensor details and CyCast’s 2D and 3D views.

Compare

Find where runs diverge

Compare a healthy run with a failing one, or examine changes after a prompt edit, LoRA adaptation or quantization. Follow differences in outputs, timing, memory and recorded model activity.

From prompt to output

  1. Prompt
  2. Tokens
  3. Layers
  4. Attention / MLP
  5. KV cache
  6. Logits / sampling
  7. Output

Available events depend on the runtime and capture depth. Every trace will identify what was captured, what is unavailable and the cost of recording it.

One capture format: .cycap

A planned portable format for AI execution. Open a saved capture without loading the original model weights, replay its events, and inspect the same evidence on another Mac.

Capture locally. Choose whether to include prompt and reply text before export, and share a file only when you decide to.

Choose what to record

Start with runtime and token information. Add layer statistics or selected attention, activations, KV-cache data and raw tensors where the runtime supports them.

Capture settings control what is saved. Display filters help you explore it afterwards.

cycap · the command-line companion

Planned

Capture from a local runtime, read a saved trace and compare runs from the terminal. Proposed command examples:

# Record a local llama.cpp run, omitting prompt and reply text
cycap capture --runtime llama-server --url http://127.0.0.1:8080 \
  --model your-model --level inference --prompts prompts.txt \
  --text omit -o run.cycap

# Inspect the saved events
cycap read run.cycap

# Compare a baseline with a changed run
cycap compare baseline.cycap candidate.cycap

CyCast Trace, the .cycap format and the cycap command are planned for a future release. These commands are design examples; the CLI is not available to install yet. No release date has been announced.

Upcoming

Next for CyCast

Features added later are sold separately.