As companies and their cloud presences scale, so does the importance of accessibility for pivotal tools and technologies, such as data integration platforms. An expanding roster of stakeholders will eventually require critical data capabilities, such as object management, creation, and configuration, across hundreds of connections and multiple teams and environments.
At Fivetran, we pride ourselves on the simplicity of our interface and the ability of our users to spin up new connections in minutes. However, even these workflows, which consist of navigating a series of menus, don’t scale very well once tens or hundreds of connections are involved.
We can make the experience of using Fivetran even easier, more convenient, and more accessible. The Fivetran REST API already enables scalable, programmatic control over Fivetran, but what if you could interact with it through natural language?
The solution lies in a Model Context Protocol (MCP) server that enables AI assistants to manage Fivetran through natural language, with approval workflows, Google Drive-backed audit trails, and complete governance controls. This is less a matter of making Fivetran simpler and more a matter of optionality and scale.
Why Model Context Protocol (MCP)?
A Model Context Protocol (MCP) is an open standard that allows large language models (LLMs) to communicate with external tools and data sources. Think of it as a plug-in for AI integrations that standardizes how any LLM discovers and uses tools without custom code for each integration. MCP offers certain advantages over simple API calls:
- Standardized AI integration: MCP provides a consistent protocol for connecting AI assistants to external systems, eliminating the need for custom API wrappers.
- Natural language operations: Users describe what they want in plain English; the AI translates intent into precise API calls with proper parameters.
- Tool-based architecture: Each Fivetran operation (create connector, check status, pause sync) becomes a discrete tool that the AI can discover, understand, and invoke with context.
- Stateful workflows: MCP servers can maintain approval queues, track request lifecycles, and persist state across sessions—critical for enterprise governance.
- Multi-user collaboration: Approval workflows become visible to entire teams of admins, enabling collaborative decision-making with complete audit trails.
These capabilities, combined with Fivetran's comprehensive API, create a powerful foundation for AI-assisted data pipeline management at scale.
The AI-assisted workflow
Building an MCP server involves designing a comprehensive workflow that strikes a balance between ease of use and enterprise controls.
1. Architecture design: Safety first
The core architectural principle is approval-aware by default. Every mutating operation (create, pause, resume, delete, modify) can be intercepted and queued for admin approval before execution. This pattern ensures that even in fast-paced demos, destructive operations can't happen without oversight. We used the decorator @requires_approval on the appropriate MCP tools to denote when the request should be queued for admin approval.
2. Google Drive integration: The audit trail foundation
The company already uses Google Drive extensively. Creating a new shared folder to facilitate approval processing and logging was the natural choice. The approval system persists all requests to a shared Google Drive file (mcp_approvals.json), providing:
- Multi-user access: Any admin can review pending requests using natural language.
- Version history: Google Drive's native versioning tracks all approval decisions.
- External review: Non-technical stakeholders can audit operations without CLI access.
3. Natural language tool design: Making complexity simple
Here's how the AI translates intent:
User: "Create a Google Sheets connector for this sales data"
AI interpretation:
1. Calls `list_destinations` to find available targets
2. Presents destination options with service account emails
3. Reminds user to share sheet with service account
4. Calls `create_connector` with inferred parameters:
- Extracts sheet URL from context
- Generates unique schema/table names with timestamp
- Sets `paused=true` for safe verification
- Defaults to ServiceAccount authenticationResult: Approval request created, visible in Google Drive and the MCP tool, ready for admin review.
4. The demo workflow (Bold prompt = Tied to an approval flow)
To test the capabilities of the MCP server, we sent it the following prompts:
Prompt 1: “What destinations do I have access to? Do I have a Snowflake dev destination?”
Prompt 2: {If the output of 1 does not show service account email} “What is the service account email for this destination?”
Prompt 3: “Create a google sheet connector with the following details
"group_id": "group_id",
"config": {
"schema": "fivetran_demo",
"table": "gsheet_demo",
"auth_type": "ServiceAccount",
"sheet_id": "12345",
"named_range": "Named_Range_1"
"kwargs": "{'paused': True, 'sync_frequency': 15, 'trust_certificates': True, 'networking_method': 'auto'}",
"approval_bypass": false”
Prompt 4: “What approvals are pending?”
Prompt 5: “Approve request xyz”
Prompt 6: “What is the status of this new connection?”
Prompt 7: “Can you sync this new connection?”Based on the results, the following worked exceptionally well:
- Natural language abstraction: Users don't need to know Fivetran API endpoints, authentication methods, or configuration schemas. "Create a connector" is enough.
- Approval workflow integration: Google Drive persistence made the approval system transparent and collaborative. Non-technical stakeholders can review pending requests without CLI access.
- Paused by default safety: All new connectors start paused, preventing accidental data syncs during demos or testing.
- Comprehensive tooling: 20+ tools covering every aspect of the connector lifecycle, from creation to migration to health monitoring.
While the following challenges emerged:
- Google Drive authentication: Setting up service account credentials and folder permissions added deployment complexity compared to local file storage.
- Error message translation: Fivetran API errors are informational; translating them into automations required simple error catalog mapping.
Fine-tuning the dynamic payload syntax took some time. This trial-and-error resulted in tools that can successfully parse error messages and offer break-fix changes where necessary. This human-in-the-loop approach made sure that response structure and syntax errors could be addressed promptly and with minimal impact on downstream users.
Once configured in Cursor, the server successfully executed the following from natural language prompts:
- Connected to Fivetran API and authenticated
- Listed destinations and derived service account emails
- Created connectors in a paused state for safe verification
- Stored approval requests in Google Drive with proper versioning
- Executed approved operations
- Provided health monitoring across connections
Lessons learned: AI data integration best practices
Although data integration through natural language processing is extremely powerful, the limitations of the technology and the inherent imprecision of language provide a few cautionary notes.
1. Design for governance from day 1
Don't bolt approval workflows onto an existing system. Building @requires_approval as a decorator from the start ensured every mutating operation could be controlled without refactoring. As teams become familiar with and usage patterns emerge, tool approval requirements can be adjusted by adding the key pattern: @requires_approval to the tool definition.
2. Persist state externally for dynamic collaboration
Local file storage works for individual developers but fails for teams. Google Drive persistence enabled:
- Multiple admins reviewing requests simultaneously
- Non-technical stakeholders auditing operations
- External compliance tools accessing the approval history
3. Make safety the default, speed the option
All connectors start paused by default. Users must explicitly resume, preventing:
- Accidental syncs during configuration
- Demo connectors running in production
- Test data polluting production warehouses
Speed-focused users can enable auto_resume in orchestrated workflows, but it's opt-in.
4. Translate technical errors into actionable guidance
Instead of raw API errors, provide recommendations.
AI-powered infrastructure management is here
This project demonstrates that AI-assisted infrastructure management isn't just theoretical; it's production-ready today:
- Natural language is the new API: Users describe intent; AI handles implementation details.
- Governance must be built in: Approval workflows aren't optional for enterprise deployments. Audit trails enable trust.
- Speed without sacrificing safety: Paused-by-default creation, approval workflows, and health monitoring protect production systems.
- Batch operations unlock scale: What took days of manual work now completes in minutes with AI assistance.
The future of data engineering isn't about replacing humans with AI but about amplifying human capabilities. An engineer who previously managed 20 connectors can now oversee hundreds with AI assistance, approval workflows, and automated health monitoring.
What's next
We will continue to evolve this MCP server with:
- Dynamic connector templates: Organization-specific best practices enforced automatically
- Health monitoring: ML-based anomaly detection for connector performance
- Self-healing workflows: Automated remediation for common failure patterns
- Cross-platform expansion: Extending the MCP pattern to other data infrastructure tools
The MCP server with Fivetran tools is currently running in production, managing 100s of connectors across multiple teams with 99.9%+ uptime. AI-assisted development didn't just accelerate delivery. It created a more secure, compliant, and maintainable solution than traditional development would have produced.
The question isn't whether AI will transform how we manage data infrastucture. Rather, how quickly will your organization adopt it? Our Professional Services team can help you answer this question, beginning with a free consultation.
[CTA_MODULE]
