Aedelgard keeps your mind separate from the model that does the thinking. Models are rented and swappable, and a new one appears almost every week. Aedelsten is the service that keeps track of them all and tells your body which to use. Every Aedelgard body asks it.
Hundreds of models from seven providers, refreshed every morning.
For machines: /v1/models · /v1/recommend (JSON, no key)
Pick a job, a budget and the providers you have keys for. Aedelsten ranks the models and says why.
Plain JSON at a public address, no key and no account: every model, its provider, context window, prices per million tokens, features and status. Any agent can read it and act on it. Responses say how long they stay fresh, so callers cache them instead of asking again.
Next to the facts sits Aedelsten’s own judgement: what a model is good and bad at, what a typical job will cost, and how it weighs one model against another. Today a person writes that advice. We are teaching Aedelsten to write it on its own key, from public sources it reads itself, with every claim dated and linked. Being built.
New models appear before anyone has priced or tested them. Aedelsten lists them as possibilities: seen, not yet recommended. Today that is the models its morning check found but could not classify. Soon it will also watch the wider web for models no provider of ours offers yet. Partly built.
When you set up Aedelgard, the models you choose from, their prices and the one marked recommended come from Aedelsten. You don’t need to know it exists. It never sees your prompts or your memory. It only answers “which model?”
One JSON API, no key: GET api.aedelgard.com/v1/models and /v1/recommend?task=…&providers=…&budget_per_1m=…. Respect ttl_seconds and cache. Prices are per 1M tokens, with cached-input prices where the provider has them. Unclassified models are listed with status: unclassified and never recommended.
Each morning a scheduled job reads every provider’s own model list. Fixed rules, not a person and not a chat model, sort each model into a family and tier and attach prices from public machine-readable catalogues and the providers’ price lists. A model that disappears from a provider’s list is marked deprecated, unless half a provider’s list vanishes at once: that is treated as their outage, not their decision. Every daily catalogue is kept as a version, and every manual change is logged with who made it and why.
We’d rather ask these in public than answer them quietly:
Tell us in the Aedelgard Discord.
Aedelsten was called Palantír until October 2026. Same service, same address, new name.
Aedelgard