Model Organizer: Choose a Provider, Then Equip the Right Models

The Model Organizer is where AcceleratedLogic AI stops being a single chatbot and becomes a workspace. The new provider-first flow makes that explicit: instead of asking which model you want, it asks which infrastructure you trust, then lets you equip models from it. This guide walks through the onboarding tabs, the three provider groups, and the privacy tradeoffs behind each choice.

The old mental model was model-first: pick a name from a long list and hope the credentials behind it work. The provider-first flow reverses that. Step one is Choose Provider: select the infrastructure that will serve your requests. Step two is Equip Models: choose which models from that provider appear in your workspace. Equipped models form your toolbox and show in the composer dropdown; the active model is the single one answering right now. You can hot-swap the active model mid-conversation without re-equipping anything.

Onboarding is organized around what you want the workspace to do, not around vendor names. The Text tab covers chat, reasoning, coding, and document work. The Image tab covers vision input and image generation, including which equipped models carry a visual badge and can accept uploaded screenshots. The Text-to-Speech tab covers read-aloud and voice output, which runs through local speech synthesis on the device. Working through the tabs in order gives you one equipped, working path per capability instead of a single default that silently fails on images or audio later.

Free access groups options with free usage. LLM7, Vireonix, and BlockRun Free can equip a hosted text model immediately without an account or API key. Puter uses sign-in. Free allowances, model availability, and rate limits depend on the provider. Browser speech uses voices available on your device.

Local providers run on your own hardware or inside the browser: built-in on-device models where supported, in-browser WebGPU runtimes, a local server such as Ollama, and lightweight in-browser transformer models you load directly. Selecting a local provider means a credentials step is replaced by a load step: confirm the runtime is reachable, pull or download the model once, then equip it. Local execution is private by construction. Nothing leaves the machine, there are no per-request charges, and offline work keeps working. The ceiling is your device: local models are smaller than frontier hosted models, the first download is large, and hard reasoning or high-fidelity vision will lag behind current hosted flagships.

API providers are direct connections you authenticate yourself: a first-party key such as Gemini, or any OpenAI-compatible or Anthropic-compatible endpoint defined by base URL, model ID, and key. The onboarding pattern is consistent: select the provider preset, paste the base URL if it is not prefilled, paste the key, fetch or enter the model list, then equip. Presets cover common clouds and local gateways, and a custom endpoint option covers self-hosted servers. Keys are stored in the browser only, never in a database the service controls, and are transmitted solely to the endpoint you configured. Verify the endpoint URL before saving, because a typo here produces authentication or model-not-found errors that look like model failures.

Every provider follows the same three beats. First, select the provider card and confirm it shows as active. Second, connect: complete the sign-in popup for free providers, load or verify reachability for local providers, or validate the key against the endpoint for API providers. Third, equip: click model cards to add them to the equipped grid, keep the set small enough to manage, then save. If connection fails, fix the layer you are on rather than equipping more models. A blocked popup, a local server without cross-origin access enabled, or a rejected key will not be fixed by choosing a different model from the same broken provider.

Onboarding can be skipped, and skipping is not an error state. Skipping leaves the workspace without an equipped model, which means chat prompts until you return to the Organizer and equip one. Skip deliberately when you are exploring the interface, setting up a machine before credentials exist, or planning to configure a local runtime later. The practical rule is simple: skip setup, not verification. When you return, equip at least one model per capability tab you care about and send a test prompt before doing consequential work.

Each group moves data differently, and that is the decision that matters. Local providers keep prompts, files, and generations on the device; they are the correct default for confidential code, personal data, and offline work. Free no-key providers minimize setup friction but maximize data movement: prompts pass through a third-party proxy to the model host, so treat them as public-side processing. Direct API providers sit in between: traffic goes to the provider you chose under that provider's terms, keys live as plaintext in browser storage, and anyone with access to the browser profile or a malicious extension on the same origin could read them. Use separate projects per client or purpose, strip credentials from prompts and uploads, use local models for the sensitive fraction of the work, and use the purge control on shared machines.

Pollinations appears under API because generation requires an API key; free allowances do not cover every model or unlimited usage. LLM7, Vireonix, and BlockRun Free require no account or key: selecting a card equips its verified default model immediately. Hugging Face remains under API and requires a token with inference permission. These categories help you discover providers; each provider still controls its own quotas, model routing, availability, and pricing.

Equip a small, deliberate set and stop. Start with a free-access or local option for quick questions, add a local model for private or offline work, and use a direct API model for long context or sustained volume. Confirm each one with a real prompt from your own workload: a code edit, a screenshot task for a visual model, and a read-aloud pass for speech. The Organizer exists to make model choice visible and reversible. Test on your work, keep what earns its place, and remove the rest.