keys, connectors and models

In ixi Local, models and pixies run on your own provider accounts. Settings is where you connect them. Add a key, connect a server ixi doesn't ship with, and choose which models your canvases offer.

Open it from the Settings button at the top of the dashboard, or from your user menu. It has three tabs: Keys, Connectors and Models. Settings exists only in ixi Local; ixi Cloud runs on ixi's keys.

keys

Keys lists every key ixi knows, grouped by what it runs:

group	keys	what they run
Model providers	Replicate · fal	canvas image, video, audio and 3D models
Pixie models	Anthropic (Claude) · Google (Gemini) · DeepSeek · xAI (Grok) · Moonshot (Kimi) · Meta AI (Muse) · OpenAI · Qwen Cloud	the LLMs pixies think with. Google's key also runs Gemini image models. OpenAI's runs GPT-6 Astra, GPT-6.1 Sol and GPT-5.6 Sol
Connectors	one per connector that names a key	that connector's requests
Other keys	anything else you saved	nothing above; sandbox scripts can still use them

To add one:

1. Press Set on its row. Get a key beside the name opens the provider's page for making one.
2. Paste the key into the box, then press Test to check it with the provider before you save. You'll see a line such as "Replicate accepted the key." or "Replicate rejected the key (401)."
3. Press Save. The row shows Set, and the next run uses the key. There's no restart.

A saved key can't be shown again. To change it, press Replace and paste the new one. The bin icon deletes it, after asking first: models on that provider stop running until you set a key again.

From environment means ixi was started with that key already in its environment. A key you set here takes its place.

which model a pixie uses

Every pixie thinks with DeepSeek V4.1 Flash by default. If you have no DeepSeek key, ixi doesn't stop the pixie. It goes down this list and uses the first row you have:

you have	the pixie runs on
a DeepSeek key	DeepSeek V4.1 Flash
an Anthropic key	Claude Sonnet 5.5
an OpenAI key	GPT-6.1 Sol
a Google key	Gemini 3.8 Flash
an xAI key	xAI Grok 4.7
a Moonshot key	Kimi K3.1 Pro
a Meta key	Meta Muse Spark 1.3
an OpenAI-compatible connector with Pixies can chat with its models on	that connector's first model

To have a pixie think with your own Claude Code instead of a key, see terminal pixies.

With nothing on the list, the pixie doesn't call any model. It answers in the chat: "No AI model is set up for pixies on this computer yet. Add an LLM key in Settings → Keys, or connect Ollama in Settings → Connectors, then send your message again." Both links open the right tab.

pin a pixie to a model

To choose a pixie's model yourself, press the sliders button in its chat header (Edit system prompt).

• Model lists every pixie model plus your connector models. A model whose provider has no key shows "— needs a key" and can't be picked.
• Thinking appears for models that take an effort level.
• Save pins that pixie to the model. Picking the default again un-pins it, so the pixie goes back to the fallback list above.
• Model recommendations (every pixie on this canvas) is the list of canvas models pixies suggest and run. Edit it to steer them toward the models you've added.

connectors

A connector is a provider of your own: a model server on this Mac, a service on your network, or any API ixi doesn't ship with. Its models show up in Models, and an OpenAI-compatible connector also gives pixies its chat models.

On Connectors, press New connector and pick a kind:

kind	for
OpenAI-compatible	Ollama, LM Studio, vLLM, OpenRouter and the like: chat models for pixies, plus image generation for canvases
Replicate-compatible	a server that speaks Replicate's predictions API
fal-compatible	a server that speaks fal's queue API
HTTP template	any JSON API: you describe the request, where the result is, and how to check on a job

Then fill in the form:

• Id (the provider name in Models): 2–32 lowercase letters, digits and dashes, starting with a letter. You can't change it later, and it can't be a built-in name such as replicate, fal or google.
• Name: what the lists show.
• Host: the server's address. A server on this Mac or your network may use http; anything else needs https.
• Key name (blank = no key): the name the key is saved under, such as OPENROUTER_API_KEY. Leave it blank for a local server that needs none.
• Auth header: how the key is sent. Left blank, the kind's usual header is used; the box shows it.

For OpenAI-compatible you can also list Models (optional, comma-separated). Left blank, ixi asks the server for its list. Set Images path only if the server's image endpoint isn't /images/generations. Two switches decide what the connector serves: Pixies can chat with its models and Canvas image models.

For HTTP template, the Mapping (request, response, async polling) box starts from a working example. {{input.x}}, {{model}} and {{jobId}} fill in values, and $.path picks a value out of a response. Leave async out for an API that answers with the result straight away.

Press Test to try the connection before you save, then Add connector. If the connector needs a key that isn't set yet, ixi switches to Keys so you can set it.

connect Ollama

1. On Connectors, press New connector and keep OpenAI-compatible.
2. Type ollama as the Id, Ollama on this Mac as the Name, and http://127.0.0.1:11434/v1 as the Host. Leave Key name blank.
3. Press Test, then Add connector.

If you have no LLM key, pixies now chat on the first model Ollama lists. Ollama has to be running for that.

paste a connector

Paste JSON takes a whole connector as one object, in the same shape the form saves:

{ "id": "ollama", "kind": "openai-compatible", "name": "Ollama", "host": "http://127.0.0.1:11434/v1" }

The other fields are keySecret (the key name), authHeader and mapping. Keys are never part of it: set the key named by keySecret in Keys.

test, edit, delete

Each connector row has Test, an edit pencil and a bin. If its key is missing, the row shows a red Set button with the key name, which takes you to Keys. Deleting a connector that models still run on asks first. Press Delete anyway and those models fail until you remove them or add the connector back.

models

Models is this Mac's model list: what a canvas model node offers, and where each run goes. The first time ixi opens, the list is filled with ixi's built-in models. If it's ever empty, Load ixi's built-in models fills it again.

To add a model:

1. Press Add model, then pick a provider, such as Replicate, fal or one of your connectors. A provider without a key shows "(no key)".
2. Search the provider, for example "flux", then press Add on a result. Or type its id under or by id, such as owner/model or llama3.2:3b, and press Fetch details.
3. Check the prefilled Name, Group, Categories (comma-separated), Output types and Description, then press Add to catalog.

A model needs no price here: it runs on your key, so no ixi credits are involved.

Routes set where a model runs, in order. Add up to four with Add route, each on a different provider. When one provider can't take the run, out of quota for example, the next one does. Left empty, the model runs only on its own provider.

The pencil on a row edits a model. The bin removes it from the model picker, after asking first. Nodes that already use it keep their settings, but they can't run it until you add it again.

limits & good to know

• A key is one line, at most 8 KB.
• A connector's Host can't hold a user name or password. Put the secret in Keys and name it in Key name.
• A model's routes: at most 4, each on a different provider.
• You can search every built-in provider. An HTTP template connector shows in search only if its mapping names its models.
• An OpenAI key runs three pixie models: GPT-6 Astra, GPT-6.1 Sol and GPT-5.6 Sol. They need no OpenAI-compatible connector. While they think, the chat shows no thinking text.
• A pixie can't use a Qwen Cloud key yet: there are no pixie models on it.
• Pixie turns on your own keys cost nothing in ixi. Your provider bills you as usual.

related

• ixi local: what runs on your Mac and what lives in ixi Cloud.
• model nodes: pick a model on a canvas and run it.
• meet the pixies: what pixies do with the model they think with.