Any model, easily
Any AI Model, Easily — Search It, Pick It, Test It
The short answer. Fred is not tied to one maker. Open any agent, type a name into the model box — "astra", "grok", "kimi" — and Fred shows what really exists: the makers you hold keys for, hundreds more through one OpenRouter key, and the models running on your own machine. Pick one, press Save & test model, and it answers you in seconds or tells you exactly why it cannot. Fred is BYOK: connect your own model access, with access and usage billed separately by your chosen provider.
The search
What really exists, not a list someone typed last spring
Most tools give you a dropdown that was current when the software shipped. Fred's box searches live: the engine's own list of models it can run right now, each maker's own list where you hold the key, OpenRouter's public catalog of hundreds of models from dozens of makers, and Ollama on your machine. Results come grouped by maker, newest first, and each one says whether it is ready, needs a key, or comes via OpenRouter, with the context size and price where known.
Under the box sit the tiles: one row per maker, the model in use first, then anything you have already tested here. Makers you have a key or a login for get six tiles, the rest three, and + N more searches that maker. Makers with nothing listed yet sit in a row of chips so you can see who else is there.
The library
Exact names, kept on your machine
Every maker spells its model ids its own way, and getting one character wrong means a blank stare from the API. So Fred keeps his own library: one file on your computer holding the exact ids for thirteen common makers — OpenAI, Anthropic, Google, xAI (Grok), DeepSeek, Moonshot (Kimi), Mistral, Groq, Perplexity, Cohere, MiniMax, Qwen and Z.AI — whether or not you hold their key yet. Where you do, the maker's own list is the truth and confirms or retires each entry. Where you do not, the spelling comes from OpenRouter's catalog by that maker's rule and is marked untested id until you test it.
The library rebuilds itself once a week, and the Refresh library link does it on demand. New models show up on their own. Nothing about it lives on our servers; plain code builds it, no model is asked.
The test
Make it answer before you rely on it
Save & test model switches the agent to your pick, asks the model for a single word, and shows the reply with the time it took — or the exact reason it failed: a missing key, an id the maker does not know, a rate limit. If a key is missing, the box to paste it appears right there. Once a model has answered on your machine it is marked tested here, and it stays that way.
Thinking
How hard each model thinks, per agent
Under the tiles Fred offers exactly the thinking levels that model has — off, low, medium, high, and on the models that support it XHigh, Adaptive or Max. Adaptive, where the model decides each turn, is a Claude 4.6 and Claude 5 feature, so Fred offers it there and nowhere else. The level with the dot is the model's own default; pick another and that is what the model gets, on the very next turn, including the chat already open. Each agent's card shows the level it runs at next to the model name, so there is no guessing.
Honest limits
- A key is still a key. The makers charge for their models on your own account; OpenRouter charges through one account for many. Local models through Ollama are free but need your hardware.
- Untested means untested. An id spelled from OpenRouter's catalog is Fred's best knowledge of what the maker calls it, not a promise; the one-click test is what turns it into a fact.
- The newest of the new. A model released this morning appears once a maker or OpenRouter lists it and Fred's weekly refresh, or your Refresh click, has run.
- Some makers publish no list. Perplexity, MiniMax, Qwen and Z.AI ids stay "untested" until you test one, because there is no list to check them against without a key.
FAQ
Questions people ask
Which AI models can Fred use?
Any model you can reach with a key. Fred keeps a local library of the exact ids for thirteen common makers — OpenAI, Anthropic, Google, xAI (Grok), DeepSeek, Moonshot (Kimi), Mistral, Groq, Perplexity, Cohere, MiniMax, Qwen and Z.AI — and one OpenRouter key reaches hundreds more from dozens of makers. Local models run through Ollama with no key at all.
Do I have to know the exact model name?
No. Type a name, a maker or a nickname — "astra", "grok", "kimi" — and the results show what exists, grouped by maker, newest first, with the context size and price where known. If you paste an id with a typo, Fred suggests the nearest real one instead of saving the typo.
How do I know a model actually works before I rely on it?
Press Save & test model. Fred switches the agent to it, asks it for a single word, and shows you the reply or the exact reason it failed — a missing key, a wrong id, a rate limit. Until then an id Fred knows but has not tried on your machine is marked untested.
How does the list stay current?
Fred rebuilds his library once a week from the makers' own lists (where you hold the key) and OpenRouter's public catalog, and the Refresh link does it on demand. New models show up on their own. The library is a file on your machine; nothing about it lives on our servers.
Can I control how hard a model thinks?
Yes, per agent and per model. Under the model tiles Fred offers exactly the thinking levels that model has — off, low, medium, high, and on the models that support it XHigh, Adaptive or Max. Adaptive, where the model decides each turn, is a Claude 4.6 and Claude 5 feature, so Fred only offers it there. The level you choose is the level the model gets, and each agent card shows it.
Is OpenRouter required?
No. Use the maker's own key for any of the thirteen common makers, or a local model with no key. OpenRouter is the backup for the obscure ones, or for anyone who would rather hold one key than many.
Before you go
Not ready to spend $99? Take the honest guide instead.
The Straight-Answers Guide to Owning Your Own AI — every hard question on these pages, answered without the sales voice. Free, by email.
Software licence: $99 once · Model-provider charges separate · 14-day refund window
Keep reading