AI providers

Every AI feature works on a self-hosted installation; you decide where the models run. There are three ways to provide them, and you can combine them.

1. Server keys (shared by everyone)#

Set provider keys in the Convex environment (deploy/compose/convex.env in the quickstart, then push it) and every member can use AI without any setup:

OPENROUTER_API_KEY=<key>   # agent, generation, critique, SVG, copy and translation
FAL_AI_API_KEY=<key>       # images, background removal, inpaint, outpaint, upscale, video

Requests are billed to your provider accounts. There are no credits on a self-hosted installation, but the per-account request caps from config/plans.json (per minute, per day and concurrent runs) still apply and protect your bill. Set spending limits in your provider dashboards as well.

Default models per task are described in providers & BYOK; exact model ids are resolved from the providers' live catalogs.

2. Members' own keys (BYOK)#

Members can add their own keys for OpenRouter, Anthropic, OpenAI, Google, Mistral, Groq, Fal or Replicate in Settings → AI providers. Keys stored on the server are encrypted with AI_KEY_ENCRYPTION_SECRET, which must be set — without it, members can only use browser-only local keys.

3. Local models#

Run models on your own hardware and set a default endpoint for everyone:

OLLAMA_BASE_URL=http://ollama:11434

With the quickstart stack, start the bundled Ollama service (profile ollama) and pull a model with tool-calling support:

docker compose --profile ollama up -d ollama
docker compose exec ollama ollama pull qwen3:32b

Because the backend runs on the same Docker network, server-run agent jobs and workflows can use http://ollama:11434 directly. For GPU acceleration, give the container access to your GPU: for NVIDIA, install the Container Toolkit and add gpus: all to the ollama service in a Compose override file.

Members can also add any OpenAI-compatible server (LM Studio, vLLM, llama.cpp, LocalAI) in their settings. See local models for model choice and CORS.

Checking providers#

Admin → Health shows which provider keys are configured, and Settings → AI providers → Test checks a specific key or endpoint and lists the models it offers. If a task has no available model (for example image generation with neither a Fal key nor a media-capable BYOK key), its buttons are disabled with an explanation instead of failing.

Without any AI#

Leave all keys unset and do not configure local models: the editor, templates, collaboration and exports work fully, and AI entry points explain how to add a provider.