Models and providers
Every model call in Labtagon AI OS goes through its LLM gateway: from the chat, from agents, from
capabilities, and from people’s own applications through /api/llm/v1. Three pages of the Control
Panel decide where those calls go: LLM providers, Models and Pipelines. Who may use a
model is decided by Model access policies (see Policies).
LLM providers, models and their keys belong to one organisation; another organisation of the same installation does not see them.
LLM providers
Section titled “LLM providers”Control Panel → LLM providers says where model calls go, and how AI OS signs in there.
- A provider Speaks one of two APIs: OpenAI-compatible (chat completions, completions, models), which most providers offer, or the System One API for typed decisions.
- It signs in with a bearer token, with a header of its own, or not at all. For well-known hosts (OpenRouter, Anthropic, OpenAI, Google, Mistral) the page says where to get a key.
- A connection is a provider’s key. With several, calls go by Priority and move on to the next when one fails (Failover: timeout, cooldown and which statuses count), each within its Budget a month.
- Keys are kept encrypted in the database and never shown again.
- Data and processing records, for each provider, where its data goes. Rule sets read it (see Policy frameworks).
Models
Section titled “Models”Control Panel → Models lists what the organisation offers:
- Models people may choose, and what New conversations start with: Auto or one model.
- Configured models: each with its provider, context, price per million tokens in and out, its use over 30 days, and when it was last checked.
- Add a model searches the providers’ lists. Use as one model offers the same model from several providers as one, or splits it again.
- Checking a model tries what it can do: Answers, Streams, Tools, Sees, PDFs, Thinks, Caches, Flex. The chat uses this, for example to give pictures only to models that see.
A model’s page has the tabs Overview, Decisions, Parameters, Data and processing, Who may use it and Calls. Prices are per provider (Context and price): they come from the providers’ price lists where there are some, and you can set them yourself. A model that a rule set withholds says Not offered: a rule set.
Pipelines: models of your own
Section titled “Pipelines: models of your own”Control Panel → Pipelines makes models of your own, named <organisation>/<name>: a pipeline of
services in front of the models that answer (ADR 0076). The services are Limits, System
prompt, Memory recall, Cache, Personal data, Tools, Routing, Guardrail,
Legal check, Keep prompts and Agent.
- Personal data in a pipeline can only be stricter than the organisation’s masking.
- Keep prompts keeps a share of the calls (1 to 100 percent) for 1 to 90 days (30 by default), masked first when masking is on. Only the organisation’s administrators read them, each time with a reason that goes into the audit log.
- Tools offers the model the tools of a profile and runs them as the caller: “headless MCP”, see AI OS from your own application.
- Agent lets an agent of the organisation answer instead of the models, if its owner allows it (ADR 0092, first stage built).
- A pipeline is published like a tool server: who approves is set by Approvals policies.
A pipeline’s page has the tabs Overview, Publish, Who may use it, Calls, Kept prompts and Versions.
Auto is a pipeline the organisation picks (setting llm.autoPipeline). With none picked, Auto is
not offered. Its routing picks a model per conversation, one that can take the turn (see
Chatting and choosing a model).
Models in your own cluster
Section titled “Models in your own cluster”Models that run inside the installation are planned (ADR 0079, proposed): an optional component of the chart with an open-source inference engine, each model offered like a provider’s. Today, models run at the providers you set up.