A native macOS application built with SwiftUI, offering visual workbenches for complex AI tasks. Train models with live charts, analyze reasoning patterns, merge models with drag-drop, and manage knowledge bases visually.
Powerful visual interfaces for complex AI tasks that would be tedious in a terminal
Complete training pipeline in a visual interface. Create datasets from your prompts, sessions, or files. Configure hyperparameters with intuitive sliders. Monitor training with real-time loss charts and deploy to Ollama when complete.
Create training data from saved prompts, chat sessions, RAG sources, or import existing datasets in Alpaca/ShareGPT format.
Visual sliders for learning rate, batch size, LoRA rank/alpha, warmup steps. Presets for common configurations.
Real-time loss curves, learning rate schedules, epoch progress, and ETA. Export charts as images.
Choose between Unsloth (fast), Axolotl (flexible), or llama.cpp for training. Auto-detects available backends.
┌─────────────────────────────────────────────────────────────────┐ │ Training Job: company-assistant-v1 │ ├─────────────────────────────────────────────────────────────────┤ │ Base Model: llama3.1:8b │ │ Training Type: LoRA │ │ Dataset: company-docs (2,847 samples) │ │ │ │ Hyperparameters: │ │ ├─ Learning Rate: ████████░░░░░░ 2e-4 │ │ ├─ Batch Size: ████░░░░░░░░░░ 4 │ │ ├─ LoRA Rank: ████████░░░░░░ 16 │ │ ├─ LoRA Alpha: ████████████░░ 32 │ │ └─ Epochs: ██████░░░░░░░░ 3 │ │ │ │ [Start Training] [Save Config] [Load Preset] │ └─────────────────────────────────────────────────────────────────┘
Train models to follow your guidelines using Direct Preference Optimization (DPO), RLHF, or Constitutional AI. Create preference datasets where you mark responses as "chosen" or "rejected", then train the model to prefer your style.
Side-by-side interface to mark responses as chosen/rejected. Import existing preference datasets or create from scratch.
Direct Preference Optimization without needing a reward model. Faster and simpler than traditional RLHF.
Define principles and let the model self-critique. Train on self-generated corrections for scalable alignment.
Create refusal datasets to train models to decline harmful requests while remaining helpful for legitimate queries.
Extract and visualize the model's internal reasoning process. See chain-of-thought steps broken down, confidence scores at each stage, and reasoning dependency graphs. Understand why the model reached its conclusions.
Parse responses to identify discrete reasoning steps, premises, and conclusions. Works with any chain-of-thought model.
See model confidence at each reasoning step. Identify weak links in the reasoning chain.
Visual DAG showing how reasoning steps depend on each other. Export as images for documentation.
Compare reasoning approaches between different models on the same problem. Find the best reasoner.
Interactive visualization of transformer architectures. Explore attention patterns, see layer activations, and understand the structure of your models. Great for learning and debugging.
Visual layer-by-layer breakdown of model structure. See embedding layers, attention blocks, FFN layers, and output heads.
Visualize which tokens attend to which. Understand how the model processes context and relationships.
See layer sizes, total parameters, memory usage. Compare architectures between models.
Export architecture visualizations as PNG/SVG for presentations and documentation.
Visual interface for creating model merge recipes. Drag models onto the canvas, adjust weights with sliders, preview merge results, and create combined models that inherit capabilities from multiple sources.
Select models from your library and drop them into the merge canvas. Intuitive visual workflow.
Fine-tune each model's contribution to the final merge. See weight distribution in real-time.
Choose SLERP (smooth), TIES (task-preserving), DARE (dropout-based), or linear averaging.
Save and reload merge recipes. Track what combinations worked best. Share recipes with team.
Test models head-to-head with the same prompts. Compare response quality, speed, and style to find the best model for your use case. Run benchmark suites and export comparison reports.
See two models respond to the same prompt simultaneously. Compare output quality in real-time.
Run standardized test prompts for coding, reasoning, writing. Get objective comparison metrics.
Response time, tokens per second, memory usage. Find the best speed/quality tradeoff.
Generate comparison reports in Markdown or PDF. Document your model selection decisions.
Manage your knowledge bases with a visual interface. Create profiles for different projects, drag-drop files to learn, search your knowledge, and preview what context gets injected into prompts.
Create separate knowledge bases for different projects. Switch profiles to change context.
Drop files, folders, or URLs onto the window to add to your knowledge base. Supports 20+ formats.
Test queries and see exactly what chunks get retrieved. Tune similarity thresholds visually.
Browse learned sources, see chunk counts, remove outdated content. Keep your knowledge fresh.
Connect to databases and explore them visually. Browse tables, view schemas, run queries with syntax highlighting, and export results. Supports SQLite, PostgreSQL, and MySQL.
Connect to SQLite files, PostgreSQL servers, or MySQL instances. Save connection profiles.
Visual tree of tables, columns, types, and relationships. Understand database structure at a glance.
SQL editor with syntax highlighting, auto-complete, and query history. Run queries and see results.
Sortable, filterable result tables. Export to CSV, JSON, or copy to clipboard.
Build custom MCP servers visually. Define tools with JSON schemas, create resources, generate TypeScript or Python server code.
Create custom agents with specific tool permissions, system prompts, and behavior rules. Test before deploying.
Develop prompts with variable templates, version history, and A/B testing. Compare prompt variants across models.
Manage Mixture-of-Experts configurations. Visualize expert routing, monitor utilization, configure selection strategies.
Configure user profiles, provide feedback on responses, train personalization adapters from your interactions.
Connect to Ollama, vLLM, llama.cpp, TGI, TensorFlow Serving, NVIDIA Triton, or any OpenAI-compatible API. Manage multiple backends from one interface.
Guided setup for new users. Configure backends, download models, set up RAG, and customize preferences step by step.
Visual task tracking for AI operations. Monitor training jobs, batch processing, and long-running tasks with progress indicators.
Deep inspection of model weights, layers, and architecture. Analyze quantization, view tensor statistics, and debug model issues.
Real-time performance metrics. Track tokens/sec, memory usage, GPU utilization, and response latencies across all backends.
Browse, download, and manage models. Search Ollama library, import GGUF files, organize by capability and size.