AI agents (MCP clients)
This page covers external AI agents — Claude, ChatGPT, Cursor, and custom clients connected via MCP.
For in-platform specialist agents (enrichment, SEO, taxonomy, approvals in the Agent Control Center), see Agentic PIM and Specialist agents.
MCP agents vs specialist agents
| Specialist agents (ACC) | MCP agents (this page) | |
|---|---|---|
| Where | Agent Control Center in PIM | External AI clients |
| Runs | Scheduled, event-triggered, missions | User-driven tool calls in chat/IDE |
| Writes | Approval queue → apply on approve | Named MCP tool calls, auditable via API |
| Quality | Insights + Catalog Quality | Same catalog; use tools to query and act |
An MCP agent is an AI client connected to CataZenta via MCP with a goal, tools, and guardrails — not a single chat message.
What agents can run
| Agent mission | Tool chain (examples) |
|---|---|
| Catalog auditor | find_products_needing_enrichment → analyze_product_completeness → create tasks |
| Publish bot | list_connections → quick_publish → get_export_job_items |
| Content writer | get_product → get_product_attributes → set_product_attribute |
| Taxonomy bot | list_categories → update_product assignments |
| Quality analyst | Agent insights and channel readiness before publish (see Catalog quality) |
Design principles
- Structured data first — agents read families and attributes, not PDFs.
- Explicit actions — every write is a named tool call, auditable via API.
- Human gates — agent proposals use the approval queue; use PIM workflows for legal/brand sign-off before
active+ publish. - Idempotent jobs — syndication via export jobs, not ad-hoc HTTP hacks.
Get started
- Agentic PIM overview — in-platform specialist agents
- MCP quickstart
- Tool reference
- Agent workflow patterns
Enterprise teams often pair agents with service accounts → recipe.