Skills, context, and tools
A skill gives the agent instructions for a task. Project context is a set of notes about your app and its files. An MCP server lets an AI client call account tools after you approve access.Separate documentation from account access
An agent can read guides and API fields to write code without accessing your accounts. A documentation MCP server gives it reference material. It does not give the agent an access token or permission to act on accounts. To run the code, put the app’s client ID and secret in its server settings. Use the values for the mode you want, such as Sandbox. Then complete the account approval flow. Start with Sandbox and check the result before using real accounts. If the agent can run commands, keep secrets in the server settings. The code still needs the correct token, permissions, and linked account ID for each request. Use an official V3 SDK for account approval, API requests, pagination, and errors.Choose the right references
Make the task explicit
For example:Check the generated integration
- Use the SDK method shown in the guide. Check its inputs and response against the API reference.
- To get, refresh, or revoke tokens, send the client ID and secret with HTTP Basic and use form fields. For Access and Vault+, send the access token and linked account ID.
- Ask for the permissions the operation needs. The token and account ID must come from the same account approval and mode, such as Sandbox.
- Read the operation’s response fields. Profiles are returned directly. Access lists and linked-account lists use different fields for pages.
- Use the right kind of ID. A linked account ID selects the creator. An OnlyFans user ID selects the fan who receives a message.
- Check webhook signatures against the original request body before using event data. Return the exact text requested by a signed verification challenge.
- If you do not know whether a change succeeded, check before sending it again. Send only one refresh request at a time for each saved refresh token.