Providers

Llooma has no models of its own. It talks to yours, through connections you add under Settings → Servers. Add as many as you like, from as many providers as you like, at the same time.

The six presets

Provider	Endpoint	Key
Ollama	http://localhost:11434	none
OpenAI	https://api.openai.com/v1	required
Claude	https://api.anthropic.com/v1	required
Infomaniak	built from your product ID	required
OpenRouter	https://openrouter.ai/api/v1	required
OpenAI-compatible	anything	optional

Everything except Ollama speaks the OpenAI chat API. For the four named providers the endpoint is worked out for you and tucked behind an Advanced disclosure, since editing it is the exception. Each preset links to its own “where do I find my API key” page.

Infomaniak asks for the AI Tools product ID from your manager, not a URL: the endpoint only varies by that number, so it is built for you. The connection cannot be synced until you give one.

OpenRouter is one key in front of several hundred models from every vendor, which makes it the cheapest way to try one before deciding whether to host it. It arrives unfiltered on purpose: no prefix means anything across a catalogue that wide, so use the model filter or the picker’s search. Drawing is not offered on it, since it serves image models through a chat-shaped route rather than the one the images page calls. Speaking is: see below.

OpenAI-compatible is the catch-all: vLLM, llama.cpp, SGLang, LM Studio, a gateway, a colleague’s box. Give it a base URL, and a key if it wants one.

Per connection

• A colour, assigned at creation from the ones not yet in use, shown wherever that connection’s models appear. With four endpoints in the picker, this is how you avoid sending a throwaway question to the expensive one.
• A label, if the provider name is not enough.
• A model filter, so a provider offering two hundred models contributes the ten you use.
• Display names, edited in a searchable sub-view: mistralai/Mistral-Small-24B-Instruct can read as whatever you call it out loud.
• Sync, one action that re-reads the model list and verifies the connection, with the date of the last successful one kept.

A stored key is shown as Key saved rather than as an empty field, so nothing looks broken, and replacing it is explicit.

Reasoning

The per-conversation reasoning toggle appears for the endpoints that can actually be asked for it:

• Ollama, which has its own native thinking mode.
• OpenAI-compatible and Infomaniak, which take the explicit chat_template_kwargs.enable_thinking flag that vLLM, llama.cpp and SGLang understand.

Hosted OpenAI and Claude reject unknown body fields, so the toggle is not offered there. Their reasoning models still reason; they simply decide it themselves.

Model parameters

Temperature, top-p, seed, stop sequences, the penalties and a token ceiling are set in Settings, Chat, and any conversation can override them from its own settings. They go to every provider now. They used to reach Ollama and nowhere else, so a temperature set on a hosted model quietly did nothing.

A second group only Ollama understands (top_k, min_p, mirostat and friends) sits in its own subsection and says so. It is not sent elsewhere on purpose: a provider that does not know a field answers 400 rather than ignoring it.

num_ctx is worth setting: it is the one case where Llooma knows the model’s real context window, which makes the load meter exact instead of measuring against the threshold you guessed.

Ollama specifics

Ollama models can be pulled from inside the app.

How Ollama loads a model, the GPU and thread counts, memory mapping and the rest, belongs to the connection rather than to a conversation, so it is set once under Advanced on the connection and every conversation held on it follows. A count left blank and a switch left on Auto both mean the same thing: Ollama decides. That matters more than it sounds. use_mmap is on by default there, so “off” and “not set” are different answers, which is why the switches have three positions rather than two.

Speech models

Sound runs in two directions, and the two are different kinds of model with lists of their own. Transcription turns what you said into text. Reading aloud turns an answer into sound. Give either one the other’s job and it answers 400, which is why the two pickers never share a list.

OpenRouter keeps both out of its main catalogue. Four hundred entries at /models, and neither the nineteen that transcribe nor the eighteen that speak are among them. Llooma asks the two narrower questions for you on every sync, so Whisper, Voxtral, Parakeet, Deepgram and Chirp turn up under transcription, and Kokoro, Orpheus, Fish Audio, MiniMax and the rest under reading aloud, with no configuration at all.

Asking those questions also settles what each model is, which no reading of a name could manage. fish-audio/s1 talks, fish-audio/transcribe-1 listens, and no substring separates them. What the provider declares beats the guess. Your own correction in Models and prices beats both.

Infomaniak answers a transcription with a job to come back for rather than with the words. Nothing about that shows in the app: you speak, and the text arrives.

Everything else is assumed to transcribe the ordinary way, /audio/transcriptions on its own root, answering with the text. Any OpenAI-compatible endpoint serving Whisper works with no entry of its own. Where the endpoint is known to take one, the spoken language goes up with the audio, and you set it in Settings, Voice. Left empty, the model works it out. It does that well on a full sentence and badly on three words.

Reading aloud takes no such default. Hardly any endpoint serves /audio/speech, so it is offered only where a provider has said it does, and a connection that has not said so has no voice models at all. Where it is offered, you pick the voice by name from a list the provider publishes, or type the name yourself when it publishes none.

In that direction the language is the voice, and every family says so differently. Kokoro’s ff_siwis is French and af_bella is American, at the front of the name. Deepgram’s aura-2-agathe-fr says it at the back. Gemini and Grok are another thing again: their voices are timbres rather than languages, and the model takes the language from the text with no field to override it.

What a call costs

Llooma tracks what an account spends, and refuses a call once a ceiling is reached. Dictation and reading aloud are now included. Before this they were neither counted nor limited.

If you have entered a price for a model, that price is used. If you have not, the figure the provider reports is.

On a gateway only the reported figure can be right. OpenRouter sends each request to whichever upstream provider it picks and bills at that provider’s rate, and it does not use one unit for everything either. Two calls, measured: Kokoro is charged per character of input, Whisper per second of audio. The catalogue calls both figures prompt and never says which is which. A price entered by hand would have been wrong by nearly four times on the first of those.

So a connection that reports shows Auto and no price fields, much as a saved API key shows Key saved instead of an empty box. There is nothing useful to type. Bill my own rate brings the fields back for one model, and it means what it says: your figure is counted instead of theirs, which is what you want when rebilling a team or working to a rate they know nothing about. Clearing it returns to Auto. Nothing shows red here either: the warning about unpriced models is about calls that would go uncounted, and none of these do.

Every other provider uses the price table, and for them it is exact. One model, one price, which is what it was built for. Infomaniak billing transcription by the minute works entirely this way.

One hole, named rather than hidden. If a provider reports nothing and you have entered nothing, the call is not counted, and an uncounted call is one your allowance never sees. It cannot be blocked after the fact, because the cost only exists once the call has been made and paid for. Entering a price is what closes it. Nothing is ever estimated for billing: the app will guess a token count to fill a progress ring, and it will not put a guess on an invoice.

Currencies are not converted. A total that mixes two is a known limitation.

How requests travel

The browser calls /api/llm/{id}, naming a connection rather than describing one, and the server resolves its address and injects the key. Keys never reach a browser, and there is no unauthenticated relay to go around it. See Security.

Yours is not in the list

Use OpenAI-compatible, which is the answer for everything the named entries are not. It works; it just cannot preset your endpoint, link to where your keys live, or know what your image models call a portrait.

Those conveniences come from one file per provider under src/lib/providers/, and adding one is a file and a line. Nothing else in the application names a provider, so contributing yours is a small change a reviewer can judge on knowing the provider rather than the codebase. And because the list is never a gate, getting a detail wrong degrades an experience instead of locking anyone out.

See Adding a provider.