Settings reference

AI providers: where to get each API key

Anthropic, OpenAI, Gemini, Grok, DeepSeek, Mistral, Groq, GLM, Llama, OpenRouter, or your own Ollama or vLLM.

Add a provider

In the portal: Admin, Settings, AI providers, then + Add a provider

Pick the provider, paste its key (or, for a server of your own, its address) and click Save. Its models then show up in every model picker under AI. A model is written as provider/model, for example openai/gpt-5.

Keys are stored encrypted and never shown again. Apps never receive them: an app asks the platform, and the platform makes the call. To stop using a provider, click Remove on its card and save.

Use a separate key for cyantula, not one you also use elsewhere. Then you can see what cyantula spends at the provider, set a spending limit there, and revoke it without breaking anything else.

Hosted providers: where to get the key

Each of these bills per token on your own account at the provider. Consoles change their menus now and then; if a step does not match, look for API keys in the account or project settings.

Anthropic (Claude)

Models appear as claude-opus-5, or in full as anthropic/claude-opus-5

  1. Sign in at console.anthropic.com and open API keys.
  2. Add credits or a payment method under Billing, or calls fail even with a valid key.
  3. Click Create key, name it after this installation, and copy it. It starts with sk-ant-api and is shown only once.

Anthropic is the default provider: a model name without a provider in front is a Claude model. With an Anthropic key in the setup wizard, every AI feature works out of the box.

OpenAI

Models appear as openai/gpt-5

  1. Sign in at platform.openai.com/api-keys.
  2. Add credit under Billing. A ChatGPT subscription does not pay for API calls.
  3. Click Create new secret key, pick the project it bills to, and copy the key (sk-…).

Google AI (Gemini)

Models appear as google/gemini-2.5-pro

  1. Open Google AI Studio, API keys with a Google account.
  2. Click Create API key and pick or create a Google Cloud project. The key starts with AIza.
  3. Turn on billing for that project when you need more than the free tier allows.

This is a Gemini API key from AI Studio, not the OAuth client you made for Google sign-in.

xAI (Grok)

Models appear as xai/grok-4

  1. Sign in at console.x.ai and add credits.
  2. Open API Keys, create a key and copy it (xai-…).

DeepSeek

Models appear as deepseek/deepseek-chat

  1. Sign in at platform.deepseek.com and top up your balance.
  2. Under API keys, create a key and copy it (sk-…).

Mistral

Models appear as mistral/mistral-large-latest

  1. Sign in at console.mistral.ai and choose a plan for your workspace.
  2. Under API keys, create a key and copy it. A new key can take a few minutes to start working.

Groq

Models appear as groq/<model>, fast hosted open models such as Llama, Qwen and GPT-OSS

  1. Sign in at console.groq.com/keys.
  2. Click Create API Key and copy it (gsk_…).

Z.ai (GLM)

Models appear as zai/glm-4.6

  1. Sign in at z.ai, open API Keys and create one.
  2. A key from the mainland China platform (open.bigmodel.cn) also works: then fill in Z.ai address (under Advanced) with https://open.bigmodel.cn/api/paas/v4.

Meta Llama API

Models appear as meta/<model>

  1. Sign in at llama.developer.meta.com. Access may need to be requested first.
  2. Under API keys, create a key and copy it.

OpenRouter

Models appear as openrouter/<vendor>/<model>, for example openrouter/qwen/qwen3-coder

  1. Sign in at openrouter.ai and buy credits.
  2. Under Keys, create a key and copy it (sk-or-…). You can give it a credit limit there.

One key reaches hundreds of models from every vendor, billed by OpenRouter.

A model server of your own

Nothing leaves your network and nothing is billed per token. Pick a model that is good at tool use if the chat or the coding agent will run on it.

Ollama

  1. Install Ollama on a machine the cyantula server can reach, and pull a model:
    ollama pull qwen3-coder:30b
  2. Ollama only listens on its own machine by default. To reach it from cyantula, start it with OLLAMA_HOST=0.0.0.0 and keep port 11434 closed to the internet.
  3. Fill in Ollama address with the address ending in /v1, for example http://gpu-box:11434/v1. No key is needed. Pulled models appear as ollama/qwen3-coder:30b.

vLLM

  1. Serve the model with tool calling on, or the chat cannot use tools. The parser depends on the model; see the vLLM tool calling docs.
    vllm serve Qwen/Qwen3-Coder-30B-A3B-Instruct --enable-auto-tool-choice --tool-call-parser qwen3_coder
  2. Fill in vLLM address, ending in /v1, for example http://gpu-box:8000/v1.
  3. Started it with --api-key? Put the same value in vLLM API key (under Advanced).

Any other OpenAI-compatible server

LM Studio, LocalAI, Together, Fireworks or a company proxy: anything that speaks the OpenAI chat API. Under Advanced, fill in its address (ending in /v1) and its key if it needs one. Its models appear as custom/<model>.

Turn on That server runs on our own hardware only for a server you run yourself. Then it is treated like Ollama: it costs nobody per token, so colleagues need no key of their own for it. Leave it off for a paid service, so colleagues bring their own key.

Personal keys: who pays for what

The keys above are the platform's own. They pay for the platform's AI: reading a repository to build it, reviewing commits, the chat and the catalog card, and apps that run on the company key.

The coding agent, and apps whose owner chose “each user's own key”, run on each colleague's own key instead. Every colleague connects theirs under Profile:

  • AI key (bring your own) takes an Anthropic key. Two kinds work. With a Claude subscription (Pro, Max or a work plan), run claude setup-token where Claude Code is signed in and paste what it prints (sk-ant-oat…); calls then count against that plan. These tokens expire: when an app starts failing for you, run the command again. Or paste an Anthropic API key (sk-ant-api…), made as described above, billed per token.
  • Keys for other AI providers takes one key per provider, made the same way as the platform's. The provider the coding agent runs on is listed first: that is the one you need today.

A model on your own hardware (Ollama, vLLM, or a custom server marked as your own) needs no personal key by default.

If a proxy in front of that server gives every colleague their own key, turn on Each colleague has their own key for it next to the server's address (for a custom server it appears once it is marked as your own hardware). The coding agent and apps that run on each user's own key then ask for a key for that server under Profile, AI key, Keys for other AI providers. The platform's own AI steps keep using the connection you filled in. Off, the platform's connection serves everybody.