Mythweaver Docs

AI connections and models

How connections, models and the links between them work, and who may add or run them.

Mythweaver talks to AI backends you choose: a local llama.cpp server, a hosted API such as OpenRouter, or anything else that speaks the OpenAI-compatible API. You set this up on two screens, Connections at /ai/providers and Models at /ai/models, and switch between them with the link in each list's header.

Three things work together:

Thing What it is Where you edit it
Connection One API endpoint: its base URL, API key, timeouts, how many replies it runs at once, and its priority. Connections at /ai/providers
Model Exactly one AI model: its friendly name, the id the backend expects, what it can do (images, audio, thinking), its context window, prices and default samplers. Models at /ai/models
Link A model reachable through a connection. A link can override the model id, carry its own weight and extra request body, and be switched off. Model connections on a connection's editor

The same model is often reachable through more than one connection: a free local box and a paid API, say. You create the model once and link it to each connection. When a reply runs, Mythweaver picks one of the model's enabled links, using the connections' priority and the links' weight. See Priority and weight.

Nothing runs until a model is linked to at least one enabled connection.

To set up your first model

  1. Open Connections at /ai/providers and click New connection.
  2. Fill in Name, Base URL and, if the backend needs one, the API key, then click Save. See Connections.
  3. Click Test connection to check the URL and key.
  4. Click Discover models to list what the connection offers, then click Add next to a model. This creates the model and links it in one step.
  5. Open the model's row under Model connections and click Send test message to make sure it answers.
  6. If you have more than one model, pick one as the Default model in Settings.

For a guided walk-through, see Connect a model.

Site and personal rows ("bring your own provider")

Every connection and model belongs either to the site or to one user.

  • A site connection or model is there for every user allowed to use the site's AI. Only site managers edit it.
  • A personal connection or model is yours alone. No other user sees it or runs it, administrators included, so nobody can read your API key. Site managers see a personal badge on their own personal rows, to tell them apart from the site's.

A link only joins a model and a connection with the same owner. A site model never runs on your personal key, and your personal model never spends the site's. The one exception is a model someone shares with you in a mode where you supply your own key: you may link it to your own connection. See Sharing and the catalog.

When you may create both kinds, a new connection or model form starts with an Owner choice: Personal (only me) or Site (every user).

To make a personal row the site's

Site managers can move their own personal connection or model to the site.

  1. Open the connection or model in its editor.
  2. Click Make site connection or Make site model.
  3. Review the confirmation. It lists what else changes: a connection takes every model linked to it along, and a model loses its links to personal connections (link it again to a site connection).
  4. Confirm. You need to have recently re-entered your password (sudo mode).

Warning

This can't be undone. A site connection or model cannot be made personal again. If the site already has a row with the same name, rename yours first.

Who may do what

Administrators grant these permissions to groups at /moderation/groups. See Users, groups and permissions.

Permission key Name What it allows
ai.connections.create Add own AI connections and models Add personal connections and models that only you see and run.
ai.site_models.use Use the site's AI models Run the site's connections and models. Without it, only your own models work.
site.ai.manage Manage site AI connections and models Create and edit the site's connections and models, and run them.
quota.ai_connections Own AI connections How many personal connections you may own.
quota.ai_models Own AI models How many personal models you may own.

The editors list what you may edit: your own rows, plus the site's when you hold site.ai.manage, plus models shared with you as an editor. Model pickers elsewhere list what you may run: your own, the site's when allowed, and models shared with you.

Restricting a site model to groups

By default, a site model is open to everyone who holds ai.site_models.use. A site manager can narrow it to some groups:

  1. Open the site model at /ai/models.
  2. Under Who may use it, below the main form, pick the groups that may run it.
  3. Click Save.

Leave the list empty for everyone allowed to use site models. Site managers can always run every site model.

A site model's Visible to users setting controls how much of it non-administrators see. See Models.

In this section

  • Connections: every connection field, testing and discovering models.
  • Models: model fields, completion modes, samplers, time limits and which model replies.
  • Priority and weight: how Mythweaver picks a connection when a model has several.