BotBound

MCP Connector Privacy Notice

This notice explains the data practices for the BotBound remote MCP server used inside Claude, ChatGPT, Cursor, and other MCP clients. It supplements the BotBound Privacy Policy. The ChatGPT specific notice remains at ChatGPT App Privacy Notice. Cursor plugin documentation is at BotBound for Cursor.

Account bound product: the BotBound MCP connector requires a BotBound account. New users create that account during OAuth so they can build their first Bot in Cursor, ChatGPT, or Claude. It creates and refines private, unlaunched Bot drafts in the connected workspace, can run a short private voice simulation or text rehearsal of those Bots, can return a privacy minimized aggregate workspace summary and per Bot next step status, and can return owner authored Training facts and rules. Confirmed Training writes land in BotBound only after the account holder enables that permission in Apps. It does not place live calls, send messages, purchase or connect phone numbers, access call recordings or transcripts, or launch a Bot from the connecting client.

Information Used to Create an Agent

When a connected user creates or updates an agent draft, BotBound processes the agent name, business or brand, intended role, goal, voice and tone preferences, operating rules, knowledge summary, intended actions, prohibited topics, and escalation preferences supplied in the conversation. This information becomes part of the user’s private BotBound workspace draft and may be edited or deleted in BotBound.

Information Returned to the Connecting Client

To fulfill a user’s request, BotBound may return the connected user’s agent draft identifier, agent name, role, business name, primary goal, tone, operating rules, prohibited topics, allowed actions, escalation policy, knowledge summary, draft lifecycle, and a link to open the draft in BotBound. Agent list results may include agent identifiers, names, roles, and channels. Agent status results may include whether the agent is on the phone network, a next step sentence, and BotBound links. They do not include the phone number itself. A coaching brief may include owner authored Training facts and operating rules, never customer communications. An aggregate workspace summary may include counts of agents and assigned numbers, recent call/text/email thread counts, unread counts, channel connection state, activation state, subscription status, last activity time, and a BotBound link. The summary does not include customer identities, phone numbers, email addresses, message bodies, transcripts, recordings, provider credentials, or secrets. For private voice, the host widget sends a WebRTC connection offer to BotBound and receives a session bound connection answer in protected metadata. The answer is not included in the model visible conversation, and no OpenAI credential is returned to the widget. A private text rehearsal returns the agent’s rehearsal reply and an explicit live_call: false flag.

Call Proposals

With the separate calls.prepare permission, your connected client can submit a Bot identifier, recipient phone number, call purpose, and instruction. BotBound stores an immutable proposal in your account and returns a review link containing only a request identifier. You must sign in to the same account, review the details and recipient permission, and select Place call in BotBound. Proposals expire for placement after 24 hours; audit records follow the general account retention and deletion policy. With calls.read, the client can read the proposal's expiry and placement status, including call and Bot identifiers, without receiving the recipient number, instruction, transcript, or recording. These permissions require explicit OAuth consent. Preparing a proposal does not place a call.

Private Voice Simulation

If a connected user selects Try private voice, the app requests microphone access through the host widget surface and creates a short lived, tool free realtime simulation for that saved draft. The simulation cannot access business systems or complete an external action. Audio processing for this optional feature is performed through OpenAI services. Do not use the simulation to provide sensitive personal information, payment card data, health information, or confidential customer records.

Account Connection and Permissions

BotBound uses OAuth 2.0 with PKCE to connect the user’s MCP client to their private BotBound account. Claude.ai connectors register dynamically. Claude Code identifies itself with Anthropic’s published Client ID Metadata Document and a loopback callback. ChatGPT uses a Client ID Metadata Document hosted on chatgpt.com. Cursor registers dynamically against Cursor operated callbacks. BotBound uses the connection only for the scopes shown during consent. The consent screen shows the connecting app name and the relying party hostname. OAuth authorization codes and bearer credentials are stored only as cryptographic hashes, are time limited, and are bound to the BotBound MCP resource.

How Data Is Used

We use connector data to provide requested first agent drafting, private simulation, owner authored Training reads and confirmed Training writes, aggregate workspace visibility, account linking, security, and abuse prevention. We do not sell or share personal information for cross context behavioral advertising. We do not use account bound agent draft inputs or private simulation audio to train BotBound models.

Retention and Controls

Agent drafts become part of the connected BotBound account and may be edited or deleted in BotBound. Aggregate workspace summaries are generated from current account state when requested and are not stored as a separate MCP record by BotBound. The general retention, access, deletion, and regional privacy rights processes are described in the BotBound Privacy Policy.

Contact

Questions or privacy requests: contact BotBound.

Last updated: October 20, 2026