Hand Claude or Codex the sales setup, keep the final say

What you can hand to Claude or Codex

MCP (Model Context Protocol) is a standard way for an AI assistant to use another product’s tools, and MyChatBot publishes several such servers, called surfaces. The Sales-Management surface covers your sales account with more than 100 tools, grouped into areas. By job, you can hand over:

• Assistants (sales_assistants): create or clone an assistant, rewrite its instructions, add skills, run private test chats (assistant_clone, assistant_update_instructions, test_chat_send).
• Channels (sales_channels): see what is connected, create a website widget, fetch setup links for the rest (channel_get_config_link).
• Clients and leads (sales_clients): read and update leads, notes, tasks, labels, pipelines and stages; export leads to a CSV or JSON file (export_clients, which returns a public download link).
• Chats (sales_conversations): read-only chats and messages, including call transcripts (get_chat_messages).
• Follow-ups and outreach (sales_outreach): draft sequences, preview an audience, run a campaign, send one message or place one call.
• Knowledge (sales_knowledge): learn your website, keep a FAQ, connect a product feed or a manual catalog.
• Orders and calendar (sales_orders, sales_calendar): read orders and bookings, and switch order taking on or off.
• Also: Meta lead-form routing (advertising), usage numbers and saved test scenarios.

Claude or Codex configures and prototypes; customer conversations keep running on MyChatBot without it.

Examples of what you could type

• “Look at my assistants, channels, pipeline and knowledge base and tell me what is missing. Change nothing until I review the plan.”
• “Add our website as a knowledge source, then open a private test chat and ask three questions a new customer would ask.”
• “List the leads in the Contacted stage that have a phone number and create a call-back task on each card.”
• “Draft a follow-up: if a lead is quiet for three days, send a short check-in. Save it as a draft only.”
• “Preview the audience for an email campaign to leads labelled warm. Show the count and a sample, and do not create it.”
• “Read the chats waiting for an operator and tell me the three questions that came up most.”

How it works

1. Connect one area. In the app open Agents → Connectors and click the small i (Connect AI tools) on a card under “From your sales platform”. You get a server address, an access key and setup for Claude Code, Codex or Cursor. One area keeps the tool list focused; the combined server is noisier for the model.
2. Inspect, read-only. Ask for an inventory and a plan; change nothing yet.
3. Draft. Let it prepare one change and show it to you first, such as the text of new instructions in the chat or a follow-up saved as a draft. Only follow-ups have a stored draft state; saved instructions apply from the next message.
4. Preview. Run a private test chat before you connect a channel or switch an assistant on (a blank assistant from assistant_create is created switched on), and an audience preview before outreach. In test chats reads run live and writes are simulated. A test shows how the assistant answers, not that a live channel delivers.
5. Approve. Review each write in your client. Codex’s writes approval mode lets Sales reads through and asks before write tools, but the docs add that message, call, delete and live-record effects need deliberate review whatever the client setting, and that reads still touch real account data.

How you stay in control

Reading is separate from writing. A client can be set to auto-approve writes, and then a successful call can reach real people, so keep the approval prompt on.

• Follow-ups start disabled. follow_up_create always stores a draft. Publishing is a separate follow_up_update, and a sequence only runs for an assistant it is attached to, which you do in the app (Toolkit → Follow ups).
• Some tools have no dry run. send_one_off_message and outbound_call act immediately. immediate_outreach_create has no draft stage: it stores a real scheduled campaign at once, and with now it can begin in about three minutes. immediate_outreach_preview_audience is read-only but reads your real customer database.
• Lead-form routing can message new leads. A lead-form mapping created in its default capture-and-outreach mode captures the lead and messages them; capture-only sends nothing. Mapping changes apply only when the call is repeated with confirm: true, and setting one needs a plan with social account connections.
• Saved changes are live. Saved configuration changes apply from the next message, chats in progress included. New assistants and widgets start switched on, and switching an assistant or channel on puts it in front of customers.
• Deletes are guarded unevenly. delete_assistant refuses an assistant with conversations unless delete_conversations is true; pipeline_delete rejects the default pipeline. client_delete also removes the lead’s chats, and deleting a follow-up, FAQ or catalog is permanent. Several have no documented guard, so your approval is the guard.
• Some actions are billed. A reply composed for a follow-up or outreach counts as one message (times your model tier’s weight), a connected voice call is billed per minute, and test chats are free.
• Keys are credentials. The access key is account-wide and shown once. Keep it, and any Product URL, out of chats, screenshots and source control.

Searching the catalog with the Product MCP

The Product surface is separate and read-only. It searches one catalog; a second catalog is a second entry. Ask in plain words, such as “find three waterproof jackets in size M and compare them”, and it can use semantic_product_search, filter_category_products, get_product_details.

There is no header, token or password: the address itself is the key. Treat the URL like a password and re-create the catalog integration if it leaks. Find it under Knowledge Base → a connected Products card → the ⋯ menu → Connect AI tools.

It cannot change products. A product feed is updated at its source, and a manual catalog is changed with update_products in sales_knowledge. See the Product MCP guide.

What this does not cover

• Hand-off and assistant working hours. The Sales tools reference documents no tool for hand-off control, stop words, an assistant’s working hours or writing into an existing chat as an operator, so set those in the app. Follow-ups carry their own working-hours schedule.
• Writing orders or calendar events. Orders and events can be read, but the reference documents no tool to create, edit or cancel them. Calendar tools need a Calendar integration.
• CRM connection. The integrations area is a read-only view with secrets stripped; no tool connects a CRM.
• Rate limits. The docs state no rate limit for Sales-Management MCP access, so this page quotes none. Individual tools do have page-size and batch caps, listed in the reference.
• Sales events as triggers. A new chat, reply, stage change or order does not start an Agents routine (a saved, repeatable workflow) by itself. The workaround is a scheduled routine that checks for changes.

Parameters: Sales tools reference. Campaigns: outreach guide. Other surfaces: MCP overview.