> ## Documentation Index
> Fetch the complete documentation index at: https://docs.datafold.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Bring Your Own LLM Provider
> Run the Datafold Migration Agent on your own LLM inference provider and credentials.
By default, the Datafold Migration Agent runs on Datafold-managed model infrastructure. With an **LLM Provider integration**, your organization can instead bring its own inference provider — your API keys, your endpoints, your cloud account — and every agent (migration goals, project defaults, the PM assistant) dispatches through it.
**NOTE**
To set up an LLM Provider, navigate to **Settings** → **Integrations** and click **Add New Integration**. LLM Provider integrations are managed by organization admins.
## How it works
* **Register a provider** with its endpoint and credentials, and select which models to expose. For most providers, Datafold fetches the available model list automatically once credentials are entered.
* **Models appear in every model picker** — goal creation, project defaults, per-role overrides, and the PM assistant — labeled with the integration name.
* **Credentials are resolved at request time.** Nothing is baked into a goal: editing an integration's key or endpoint takes effect on the next agent request, without recreating goals.
* **Credentials are encrypted at rest** and never returned by the API — secret fields are masked on read and kept unchanged when a masked value is re-submitted.
* **Costs are tracked automatically.** Token usage and estimated spend per model appear in goal cost breakdowns and count toward goal budgets.
## Supported providers
Anything that speaks the OpenAI `/chat/completions` protocol — Together, Fireworks, Groq, OpenRouter, Mistral, vLLM, LM Studio, Ollama, and most other inference services — works through the universal **OpenAI-compatible** provider.
## Common configuration
Every provider shares a few fields:
| Field | Description |
| --------------------- | --------------------------------------------------------------------------------------------------------------------------------------------- |
| **Name** | A display name for the integration (e.g. `Azure Prod`). Shown next to its models in pickers. |
| **Models** | The model ids this integration exposes. Fetched automatically where the provider supports listing; models can always be added by id manually. |
| **Max Output Tokens** | Per-response output cap. Pre-filled with a family-appropriate default; override if your deployment differs. |
| **Context Window** | Model context size in tokens. Pre-filled from Datafold's model metadata; drives conversation management for long agent runs. |
## Model quality matters
Migration agents lean heavily on tool calling and long-context reasoning. Small or non-tool-calling models will technically dispatch but perform poorly on agentic work. We recommend frontier-class models (Claude Sonnet/Opus, GPT-5 family, Gemini Pro/Flash, Mistral Large) for goal execution.