Skip to main content

Use Only Your Own API Keys (BYOK)

With Only your own API keys (BYOK), every AI request of your organization reaches the provider on your organization's API keys, never on MINEO's: conversations, Work, voice and MINEO's AI helpers. Use it when your data processing agreements or your billing require that no request goes through MINEO's provider accounts. Off, your organization uses MINEO's models: there is no mix of both.

Everything is configured by organization administrators in Organization → AI → Models. Your pricing plan must allow BYOK models.

Turn it on​

  1. In AI → Models → General, switch on Only your own API keys (BYOK).
  2. Click Save. Before anything is saved, MINEO lists what changes:
    • the default models that are MINEO's, which are cleared — the organization's and each region's;
    • the features that turn off, because nothing of yours serves them yet;
    • the assistants that switch to one of their own models, or stop answering until they have one — per project, and once for the plan's assistants, which are the same in every project.
  3. Confirm. The Credentials, Models, Groups, Deployments, Defaults and Effective access tabs appear, and the checklist under the switch shows what is still missing. It can be turned on before any model of yours exists.
Until your models are set up, your organization has no AI

Turning it on takes MINEO's models away at once. Set up your models right after, following the steps below, or turn it on when nobody is using AI.

Set up your models​

The checklist under the switch (AI → Models → General) ticks each step as you complete it.

  1. Add a credential in Models → Credentials: a name, your API key and the providers it serves (OpenAI, Anthropic, Google; a LiteLLM proxy may serve all three with one key, and needs its Base URL). Use Test connection before saving: it tries the key with each provider. Through a proxy, Claude and Gemini models get no file storage of their provider: a large attachment or a spreadsheet is not placed in the code execution environment (the model reads a partial view of a spreadsheet instead), and documents travel with every request. To have Claude work on files with code, deploy the model with a credential that holds a direct Anthropic key (no Base URL). OpenAI models keep their files through a LiteLLM proxy that has file storage configured (files_settings).
  2. Create your models in Models → Models with +: pick a base model of the catalog and a name. Saving opens the model's page, where its Regions tab sets where it runs and with which model ID (see One model in several regions).
  3. Deploy them: in the model's Deployments tab, add who may use it and with which credential (see Deployments). A model with no deployment is offered to no one. To give several models at once, put them in a group under Models → Groups and deploy the group; a new deployment offers a group first. Check it works with Test model, in the model's ⋯ menu in the list or on its page: it sends a minimal request with the key of its most specific deployment and shows the reply, how long it took and whether MINEO has a price for it. An image model makes a real, low-quality generation. Models → Effective access shows what a given member gets in a given project.
  4. Pick your defaults in Models → Defaults, in the Organization column (and, with regions on, in each member region's column if it uses other models):
    • Model for MINEO's AI helpers — writes instructions, starters, form fields, assistant setups, skills, Dockerfiles and code completions; the notebook assistant runs on it too.
    • Dictation model and Reading aloud model — OpenAI speech models, created in step 2 from a speech base model and deployed with an OpenAI credential.
    • Compaction model and Image generation model — optional.
  5. Review your assistants: an assistant whose allowed models are all MINEO's will have nothing to run on.

What changes for your users​

FeatureWith your own modelWithout one
Conversations and WorkRun on the assistant's own modelsThe composer says the assistant has no model available and can't send
MINEO's AI helpers (magic wands, code completion)Use the model for MINEO's AI helpersThe wands are disabled and their tooltip says why; code completion stops asking
Dictation and reading aloudUse your speech modelsThe buttons are disabled and their tooltip says why
Notebook assistantUses the model for MINEO's AI helpers, with its keyDisabled, and its tooltip says why
Assistants are never rewritten

Taking MINEO's models away doesn't change any assistant's configuration. Each conversation uses the assistant's default model if it is allowed, otherwise the first of its allowed models that is. Turning the switch off again brings MINEO's models back as they were, and your own stop being offered (they are kept, for when it is on again). Meanwhile an assistant on one of your models runs on MINEO's model of the same base model, if your plan offers it, and default models that are yours go back to MINEO's — the organization's and each region's.

Deployments: who uses each model and with which key​

A deployment gives a model, or every model of a group, to some people, and they call it with the deployment's credential: what they spend goes to that key. Each one has:

  • It gives: a group of models or one model. The pickers have Manage, New and Reload at their foot: what you create there is picked at once.
  • Credential: only the ones that serve the providers of its models are offered.
  • For whom: the whole organization, or projects, groups of members or users. Picking several creates one independent deployment for each.
  • Order: breaks a tie between two deployments just as specific. A new one goes last: 10 more than the highest of the table it is added from.

A deployment gives its models in every region they run in (see One model in several regions).

When several deployments give a model to the same person, the most specific one wins: one for a user before one for a group, a group's before a project's, a project's before the whole organization's; then the lowest order. So a department's key can serve its group everywhere while a project with its own budget calls with its own key.

The deployments are listed on the model's page (its own and its groups'), on the group's page, on the credential's page and, all of them, in Models → Deployments. In each table they can be edited, deleted — several at once, selecting them — and moved one above or below the next with the arrows of the Order column.

Department setupDeployment
One key for everyoneThe model (or the group), the key, nothing else
One key per departmentOne per department's group, with its key
One key per projectOne per project, with its key
A person who pays with their own keyOne for that user

One model in several regions​

When your provider serves the same model in several regions under a different ID (for example a LiteLLM proxy with eu-gpt-5.6-luna, us-gpt-5.6-luna and in.openai.gpt-5.6-luna), add it once. On the model's page, the Regions tab has a numbered row per region with the ID it is called with there (empty: the base model's): add one with the + button at the top right, remove and reorder them.

  • With regions on, a member uses the row of their AI region (see Models by region), in the order of their regions; one without a region, or anyone with regions off, uses the first row.
  • A region's May use models of decides which rows its members may use: a model is offered while one of its regions is allowed.
  • A deployment gives the model in all of its regions: the member calls the row of their region with the deployment's key.

Retire a model​

A model's Lifecycle tab shows when its provider retires it — its base model's date, kept up to date from LiteLLM's model catalog — and a warning as the date nears.

To move everyone from one model to another (a new version, a cheaper one), set the old one's Replaced by there, or add the old one to the new one's Replaces: whatever used it — assistants, default models, conversations — runs on the substitute from then on, without editing anything else, and the list counts it as Retired. Clear it to go back. The model pickers show Replaced by on a replaced model.

Next steps​