Build an agent with your AI client

With the MCP server connected, your own AI assistant (Claude, ChatGPT, Cursor and others) can build a SendSeven Conversational Agent for you. The assistant writes the content; SendSeven stores, tests and runs it. Nothing here uses the paid Agent Builder API, so only the test runs of your agent use AI credits.

Before you start

1. Connect the MCP server and keep the Agent Builder, Knowledge Base Authoring, Contact Data Model and Flows capabilities enabled (they are on by default). Connections made before these capabilities existed must be reconnected once: remove SendSeven in your AI client, add it again and approve the new permissions.
2. Your plan must include Conversational Agents (Professional or higher). Otherwise tools answer plan_upgrade_required.
3. Ask the assistant to run check_permissions if you are unsure what the connection may do.

The guided way

Ask your assistant to use the build_conversational_agent prompt. It interviews you step by step:

1. Your business, website and audience
2. Goals and what the agent must never do
3. Persona, tone and languages
4. Knowledge: it crawls your website (crawl_website), adds documents (upsert_knowledge_base_document) and FAQ entries (upsert_faq_entries), and shows the indexing status (list_knowledge_base_documents)
5. Contact fields and tags the agent may read or fill (manage_custom_fields, create_tag)
6. Skills with goals (upsert_skill)
7. Hand-off rules to your team
8. A test of the draft (test_agent), which reports the AI credits used
9. A readiness check (get_agent_readiness)
10. Publishing and attaching it to a channel, after your explicit yes (publish_agent, attach_agent_to_channel)

Two more prompts help with parts of the job: write_skill for a single skill and design_agent_flow for an automation flow around the agent.

Safety defaults

• create_agent always creates an unpublished, inactive draft, and reuses an agent with the same name instead of creating a duplicate.
• Edits go to the draft, never to the live agent. If the draft changed in the meantime, the tool reads it again and retries once; a second conflict returns draft_conflict.
• publish_agent and publish_flow change what customers see. They only run with confirm=true, which the assistant sets after you agree.
• Attaching an agent to a channel needs publish permission; without it the tool returns an error that explains this.

Reference resources

The server exposes four guides as MCP resources the assistant can read: sendseven://guides/flow-format-spec, sendseven://guides/skills-and-goals, sendseven://guides/agent-fields and sendseven://guides/run-assistant-node. They mirror these pages:

• Flow format specification
• Skill goals
• Agents API and Skills API

What happens behind the tools

Step	API guide
Draft, revisions, publish	Drafts and publishing
Readiness check	Readiness
Test turns and credits	Test console and turns
Flows	Flows overview

When a permission is missing

If a call needs a permission your connection does not have, the tool returns permission_missing with the scope and capability named. Reconnect SendSeven in your AI client and approve it. The underlying API reports the same situation as missing_scope with reason: token_scope. If the tool returns forbidden, your workspace role does not allow the action, and a workspace admin has to change that; reconnecting does not help.