OnDevice vs Agent Lab

OnDevice is a local-first AI studio for iPhone. Run compatible language, vision, speech, and image models on your device. Browse models and choose what to download. Model size, memory use, and speed depend on your hardware. CHAT AND CREATE Ask questions, write, work with code, save snippets, and export conversations. Use camera or photo input with a compatible vision model, or use built-in OCR without downloading a vision model. Voice conversations use local Apple speech by default, with optional downloadable neural voices. Compatible image models support local image generation, which can be restricted with a parental PIN. YOU CHOOSE THE CONNECTIONS Local inference uses installed models. Network features are optional: model discovery and downloads, web search, paired Mac tools, a local API, and iCloud sync. Web providers may receive the queries or content you choose to send and may require their own API account. No app account is required for local use; there is no analytics SDK. REQUIREMENTS Requires iOS 18 or later. Downloadable models need storage and sufficient available memory. The app may pause or refuse heavy workloads when the device is too hot or memory is low. Model compatibility and performance vary by device.

Agent Lab is a learning and experimentation playground for AI agents, built for AI learners, agent developers, and product managers. It doesn't just run agents — it makes every step of an agent's execution visible, so you can finally understand how agents actually work. Unlike an ordinary AI chat app, Agent Lab is about learning, experimenting, debugging, and visualization. On every run you can see the prompts, tool calls, retrieval, tokens, and latency — nothing is hidden. Get started fast · On-device local models: download a small Qwen model on first use (Wi-Fi recommended), then run it fully offline — no API key and no second machine required. Cost is always 0, and tokens and latency are shown for real. · Download on demand: grab other open-source models your device can actually run, with capability gating based on your chip and memory. · Cloud models: works with Chat Completions-compatible APIs (DeepSeek, Moonshot, OpenRouter, and more) and Anthropic Claude — just enter your key, base URL, and model. Three built-in demos · Customer Support Agent — demonstrates tool calling and workflow with a built-in mock order system, no real backend needed. · RAG Knowledge Base — comes with curated reference docs (LangGraph, MCP, LLM APIs, Python) and lets you upload your own PDF / Markdown / TXT, then inspect the retrieved chunks, similarity scores, and final answer. · Research Agent — shows how an agent plans on its own: Planner → Search → Read → Summarize → Answer. Unified visual debugging Every demo offers three views: · Timeline — the full step-by-step sequence of LLM and tool calls · Graph — the workflow graph, highlighting the active node as it runs · Console — expand any step to see System / User / Assistant prompts, the complete LLM request and response JSON, tool requests and responses, tokens, estimated cost, latency, and errors Security and privacy · API keys are stored encrypted on-device via the system Keychain — local only, never uploaded to any server. · Uploaded knowledge-base files and the vector indexes built from them stay on your device and can be deleted anytime. · On-device inference runs fully offline; your data never leaves the device. No sign-up, no login — open it and go. Make every step an agent takes visible.