1. The Big Idea — One Format to Rule Them All
Almost every AI server today speaks the exact same standard protocol: OpenAI-compatible REST API.
Whether it's Ollama on your laptop, LM Studio on your desktop, Groq's ultra-fast cloud, OpenRouter with 200+ models behind one key, Mistral, DeepSeek API, or local vLLM / LocalAI instances — they all answer at an HTTP endpoint ending in /v1/chat/completions and expect:
Add custom endpoints to Model Organizer and select their models for chat. Support for tools, images, and other features depends on the endpoint and model.
2. Ollama — Run Models on Your Own Hardware
What it is: Ollama runs a model on the machine hosting the Ollama service. When the endpoint is local, the prompt is processed there instead of by a hosted model provider. Download the app and model first; network access may still be used by this site and by any other connected provider or feature.
- Download and install from ollama.com (Mac, Windows, Linux).
- Open your terminal and pull a model:
ollama pull llama3.2ollama pull qwen2.5-coder:7bollama pull llava
Browsers prevent web applications from connecting to local ports by default. If you run default ollama serve without CORS enabled, Accelerated Logic AI will show "Disconnected".
Start Ollama with CORS origins enabled:
If socket in use: Right-click Ollama llama icon in system tray -> Quit Ollama, then run command again.
Verify: Open http://127.0.0.1:11434 in your browser. It should say "Ollama is running".
- Open Model Organizer -> Select Provider Ollama.
- Click View Local Models. Connected indicator turns green when detected.
- Your installed models (e.g.
qwen2.5-coder:7b) appear in the list. Click Equip. - You can also type any model name and click Pull to download directly within Accelerated Logic AI with real-time progress indicators.
3. LM Studio — Desktop GUI for Local Models
LM Studio offers a rich visual desktop interface for running GGUF local models.
- Download LM Studio and load a model (e.g. Llama 3.2 3B Instruct).
- Go to the Local Server tab on the left.
- Ensure CORS is enabled in settings (toggle Enable CORS on).
- Click Start Server (default runs at
http://localhost:1234/v1). - In Accelerated Logic AI: Go to Model Organizer -> OpenAI Compatible -> Base URL:
http://localhost:1234/v1-> enter exact Model ID -> Add.
4. OpenRouter — One Key for 200+ Frontier Models
OpenRouter provides unified proxy routing to Claude, GPT-4o, Llama 405B, DeepSeek, Qwen, and Mistral through a single API key.
- Create an API key at openrouter.ai.
- In Accelerated Logic AI: Open Model Organizer -> OpenAI Compatible -> select provider preset OpenRouter.
- Base URL auto-fills to
https://openrouter.ai/api/v1. - Paste your API key (
sk-or-v1-...). - Click Fetch Models to dynamically load 200+ available models in a dropdown, or enter Model IDs like
anthropic/claude-3.5-sonnetordeepseek/deepseek-r1.
5. Groq — Ultra-Fast Hardware Inference
Groq delivers high-speed inference (300-800 tokens/sec) on Llama and Qwen open models.
- Get a free key from console.groq.com (
gsk_...). - Select preset Groq in Model Organizer (Base URL:
https://api.groq.com/openai/v1). - Enter Model IDs such as
llama-3.3-70b-versatileorqwen/qwen3-32band save.
6. Other Compatible Endpoints & URL Cleaning
- Together AI: Base URL
https://api.together.xyz/v1 - Mistral API: Base URL
https://api.mistral.ai/v1 - DeepSeek API: Base URL
https://api.deepseek.com/v1 - Custom vLLM / LocalAI: Select Custom Endpoint, enter your URL (e.g.
http://192.168.1.10:8000/v1).
💡 Automatic URL Cleaning: If you accidentally paste a full endpoint like .../v1/chat/completions, Accelerated Logic AI automatically strips redundant path segments and trailing slashes.
7. Managing Custom Models (Equip, Duplicate, Edit, Delete)
All custom providers appear in the Configured Custom Models manager.
Clicking Equip adds the model to your active toolbox and sets it as active in your composer.
Duplicate any custom endpoint configuration to quickly adjust provider keys or test secondary endpoints.
8. Privacy & Cost Comparison
| Provider Type | Privacy Level | Cost | Best Used For |
|---|---|---|---|
| Ollama / LM Studio | Local inference when the configured endpoint is on-device; setup and other app features may use the network | No hosted inference fee; hardware, electricity, and model license terms still apply | Confidential code, offline work |
| Groq / Mistral | Encrypted Cloud Transit | Pay per token / Free tier | Ultra-fast agent loops |
| OpenRouter | Unified Router Proxy | Unified Account Billing | Model comparison & rare models |
9. Mixed Content & CORS Gotchas — Troubleshooting
Hosted HTTPS apps cannot directly fetch unencrypted http://localhost endpoints in some browser security policies.
Fix: Enable CORS properly, run local development server with npm run dev, or use cloud endpoints (Groq/OpenRouter) when on hosted previews.
Server rejected request origin header.
Fix: For Ollama, launch with OLLAMA_ORIGINS="*" ollama serve. For LM Studio, turn on "Enable CORS".
Ensure API key has no trailing spaces. For model IDs, verify exact tag (e.g. qwen2.5-coder:7b) via ollama list.
10. Practical Power-User Stack & Fallback Chains
ollama:qwen2.5-coder:7b— Private local codinggroq:llama-3.3-70b— High-speed general assistantopenrouter:claude-3.5-sonnet— Complex reasoning when stuckputer:deepseek-v4-flash— Free backup
In Settings -> Model Organizer you can save a preferred order of alternate models for an equipped model (a fallback chain).
Important: the saved order is configuration data only — request execution does not currently read it to retry a failed request. If the primary model fails, the error is reported; pick another equipped model and retry manually. Do not rely on automatic failover (see Part 8).
11. Choosing Between OpenRouter and a Direct Provider Key
Both routes speak the same OpenAI-compatible protocol, so the choice is about economics and control, not compatibility:
| Consideration | OpenRouter (aggregated) | Direct provider key (Groq, DeepSeek, Mistral...) |
|---|---|---|
| One key, many models | Yes — swap models without new accounts | No — one account per provider |
| Pricing | Pass-through pricing plus OpenRouter's own fee structure; check their current pricing page | Provider-native pricing, often cheaper at volume or via free tiers |
| Model coverage | Broadest — including models you cannot get directly otherwise | Only that provider's catalog |
| Privacy posture | Prompts transit OpenRouter's relay before reaching the upstream lab | One hop: browser (or proxy) → provider |
| Best for | Exploration, rare models, side-by-side comparison | A model you already know you want, high volume, tightest latency |
A sensible progression: AL Router / Puter to find the model you like → OpenRouter to evaluate it cheaply → direct key (or local Ollama) once you commit.
12. LAN & Home-Network Endpoints — Reaching a Model on Another Machine
Running the model on a desktop and the chat on a laptop is a common, supported setup — with two gotchas:
Binding the server to the LAN. Ollama listens on localhost by default. Set OLLAMA_HOST=0.0.0.0:11434 alongside OLLAMA_ORIGINS="*" before ollama serve so other devices can reach it, then use http://192.168.x.x:11434/v1 as the Base URL.
Firewall and router rules. Windows Defender Firewall must allow inbound connections to the Ollama process on port 11434, or the LAN device will see a timeout rather than a CORS error — a different failure signature. Verify from the second machine with curl http://192.168.x.x:11434/api/tags (or open the URL in its browser) before debugging the app.
Never port-forward raw Ollama to the internet. An unauthenticated 0.0.0.0 listener on a public IP is an open invitation. LAN-only is fine; internet exposure needs a reverse proxy with authentication in front, and even then is discouraged.
Mixed-content nuance. If the app is served over HTTPS and your endpoint is plain HTTP on a LAN address, some browsers block the request as mixed content or a private-network-access violation. Prefer serving your own dev build locally (or use a trusted certificate) when testing cross-device setups.
13. Verifying an Endpoint End-to-End with curl — Before Touching the App
When an endpoint misbehaves, a thirty-second curl test splits "server broken" from "app config broken" instantly:
# 1. Health check (Ollama)
curl http://127.0.0.1:11434/
# expect: "Ollama is running"
# 2. List installed models
curl http://127.0.0.1:11434/api/tags
# expect: JSON with a "models" array; note exact names
# 3. One real completion
curl http://127.0.0.1:11434/v1/chat/completions -H "Content-Type: application/json" -d '{"model":"qwen2.5-coder:7b","messages":[{"role":"user","content":"ping"}]}'
# expect: JSON with choices[0].message.content
Server is fine — the problem is app-side: wrong Base URL, missing key field, or the app fetching from a different origin. Compare the exact URL the app calls.
Read it — most OpenAI-compatible servers return a JSON "error" object with a message ("model not found", "out of memory"). That message is the real diagnosis.
Firewall, VPN, or the model is still loading into memory (large models can take minutes to first token). Watch the server console for load progress.
14. Using the App on a Shared or Public Computer
The credential model (Part 22) assumes your own browser profile. On a shared machine, adjust accordingly:
- Use a private/incognito window. Site data dies with the window, so nothing — chats, keys, tokens — persists for the next user. You will re-equip models each session; that is the point.
- Prefer session-scoped keys. If you paste a provider key, remember it lives in browser storage for the life of the profile. Private mode keeps that blast radius to zero after you close the window.
- Never connect GitHub or Google Drive on a shared machine — those credentials should not outlive the session (GitHub PATs live in persistent storage; Drive tokens are session-scoped but accounts stay connected).
- Signing out properly: clear site data for the app in browser settings, or use the in-app Clear/Purge controls (Part 17), and close the window.