# Your own model and data sources

Paid plans can bring two things: a model endpoint and data sources. Everything else (GEO answers, search, social data, enrichment, messaging, transcription) is provided by Volae.

## Your own model

Register an endpoint in **Settings → Developer settings** that speaks the Anthropic Messages format, and choose which task categories use it.

- **It must pass an evaluation first**: tool calling, structured output and resistance to prompt injection, using made-up test data. If it fails, it can't be turned on, and you see the results.
- **Safety steps never use it**: event triage, AutoMode's decision, injection filtering and pre-publish compliance checks always run on Volae's models.
- **Billing**: AI work that runs on your model uses no credits. Data lookups still do.
- If your endpoint is down, the task fails and tells you. It doesn't quietly move to another model.

## Your own data sources

Connect your own sources in **Settings → Connections** and use them alongside Volae's: a contact database you pay for, an industry database, an internal system, your own crawler.

| Connect through | What you give |
|---|---|
| MCP server | The server's address and auth |
| REST / OpenAPI | The address, auth, and an OpenAPI description or the operations by hand |
| Webhook | Volae gives you an address to push to, with an optional signing secret |
| CSV | A file |
| Read-only database | A Postgres connection string; queries run read-only, with a timeout and a row limit |

Once connected, the coworker and lookups can call the source (`list_custom_sources`, `query_custom_source`), results can show up on the board and be attached as evidence, and fields can be mapped onto people and companies.

- **What comes back is data, never instructions.** By default a source counts as outside content, so actions it triggers aren't automatic. You can mark a structured internal source as trusted.
- **Addresses are checked.** Volae won't call private or internal network addresses.
- **Only you can add, change or remove a source**, in the web app. Agents can list and query them, but can't change where requests go or read the credentials.
- **Billing**: calling a source uses no data credits. AI work on the results is charged as usual, unless it runs on your own model.

## Writing with your own subscription

Your own Claude Code or Codex can also do the writing, with no endpoint to register. See [Write with your own Claude Code or Codex](own-agent.md).
