# Set up the agent

> The agent has the same capabilities whichever path you choose — what differs is where the model runs and how you authenticate. Pick BYOK for provider control, built-in for zero setup, or MCP to drive Recense from Claude or ChatGPT.

*Source:* https://recense.ai/docs/ai-overview

## The three paths

- **BYOK** — Your API key, calls go directly from your browser to the provider. Best for individual users who want provider and model control.
- **Hosted** — Recense-managed agent, no key required. Best for teams who want zero-setup AI with usage metering.
- **MCP** — External assistant (Claude Code, Claude Desktop, ChatGPT) controls your session. Best for users who want AI assistant integration from their desktop app or IDE.

## What the agent can do

- Describe the dataset structure, variables, and metadata.
- Build and modify tables on the canvas.
- Apply filters, weights, and measures.
- Interpret results and explain statistical output.
- Assist with text coding workflows.

## One behaviour model

The in-app agent and MCP tools act on the same session. Prompts, tools, and visible state stay aligned regardless of which path you use.

## Before you start

- Sign in so your plan and access level can be resolved.
- Open a dataset before expecting useful AI output.
- For MCP workflows, keep the Recense browser tab open — the AI assistant needs to connect to your live session.

## Detailed setup

- BYOK setup — add your own Anthropic, OpenAI, Google, or Fireworks key.
- Hosted agent — use the built-in agent with no key management.
- MCP setup — connect Claude Code, Claude Desktop, ChatGPT, or custom clients.

## Next steps

- **[Agent recipes](/docs/agent-recipes)** — Prompt patterns that get reliable analysis out of the agent.
- **[Write dataset instructions](/docs/dataset-instructions)** — Give the agent persistent context about your survey.
- **[Bring your own keys (BYOK)](/docs/bring-your-own-keys)** — Connect your own Anthropic, OpenAI, Google, or Fireworks key.