Custom LLM providers
Point an employee at any OpenAI-compatible endpoint, your own server or a model from OpenRouter.
Beyond the built-in model tiers, you can connect your own model. Anything that speaks the OpenAI API works: a hosted provider, a server you run yourself, or a model picked from OpenRouter.
See every model you run
LLM Providers lists every endpoint you have connected, with its model name and its address.

Each provider can be switched on or off, edited or removed, and one of them is marked as the default. The last block, Model for every employee, sets that default: it applies to every employee that has not been given its own model.
Add a model by hand
Four fields and it is connected: a name, the base URL of any OpenAI-compatible endpoint, an auth token, and the model name.

Ollama on your own hardware works the same way, with a local address such as
http://10.0.0.14:11434/v1. Tokens are encrypted at rest.
Or take one from OpenRouter
Add from OpenRouter opens the catalog: search, sort by price or context size, and see what a million tokens costs before committing.

Pick a model and it is added as a provider, ready to assign.
Per-employee assignment
A provider is available to the workspace; the choice is made per employee.
Open Configure employee → Intelligence. Custom providers appear under a
Custom Providers heading, below the built-in engines, and the one in use is
ticked.
Two employees in the same workspace can therefore run on entirely different models: a heavy one for the work that needs it, a cheap one for high-volume replies, or a local one for anything that must not leave your network.
Choose what its employees may do
On each custom LLM in Settings → LLM providers, a shield button opens its permissions: no restriction, one of three presets, or your own rules.
| Preset | What it does |
|---|---|
| Open | No command is blocked. |
| Standard | Blocks risky commands: sudo, ssh, scp, rsync, docker, git push, npm publish, crontab, shutdown, reboot. npm install, pip install, curl and git still run. |
| Strict | Only reading and writing files, the employee's skills and a few safe commands (cd, ls, cat...). Everything else (npm, pip, curl...) is blocked. |
Your own rules name a tool (bash, read, write, edit, grep,
find, ls, or any tool), an optional pattern (the command for bash, such as
npm *; the path for file tools) and an action: Allow, Ask or
Deny. A default action covers everything else. Deny wins over Ask, Ask over
Allow.
- A blocked action never runs, whatever the model insists: the employee is
told it is not allowed. Chained or wrapped commands are checked too
(
cd x && npm i,bash -c "npm i",timeout 30 npm ci...). - Ask shows a card in the employee's chat with the command, Allow or Refuse. Anyone who can open the chat may answer; without an answer, the action is refused after 15 minutes.
- Test a command: type one (
npm install express) and see at once whether it is allowed, asked or blocked, and by which rule. - Journal: the latest actions blocked, approved, refused or left without an answer, with the employee, the date and who decided.
Changes apply from the employee's next turn.
Network access
A globe button next to the shield sets the LLM's network access, apart from
the command rules: no restriction, block these IPs, or only these
IPs. Lists take IPv4 addresses and ranges such as 10.0.0.0/8, one per line;
an empty only these IPs list means no internet at all. Name resolution (DNS)
and Geta.Team itself always stay reachable. The rule also covers the server
browser of these employees.
Together, the two buttons let you run a model you trust less, a local one or a cheap one, inside a fence you chose.