AI4Love MCP Integration — How It Works

AI4Love provides a Model Context Protocol (MCP) server that lets AI assistants — Claude, ChatGPT, or any MCP-compatible client — query your supporter data conversationally. Staff ask questions in plain language and get answers grounded in live data. All 19 MCP tools are strictly read-only.

How MCP Works

A staff member asks a question in their AI assistant (e.g., "Who should we call this week?")
The assistant calls one of 19 read-only MCP tools
The MCP server validates the user's access key against Airtable
If valid, the server queries your organization's dedicated Airtable base
Results pass through allow-list field filtering
Filtered data returns to the assistant, which formats the response

The assistant never accesses your data directly. Every request goes through the MCP server's validation, filtering, and rate-limiting layers.

What the Assistant Can Do

All 19 tools are read-only:

Supporter queries: Look up individuals, filter by status or giving level, view full relationship timelines
Activity queries: Donations, volunteer hours, event attendance, communication history
Insight queries: AI-generated recommendations, at-risk supporters, conversion opportunities, recognition milestones, campaign recommendations, relationship alerts
Research queries: Search nonprofit sector research and your organization's internal knowledge base
Export: Structured data export for external platforms (with first/last name split for CRM compatibility)

What the Assistant Cannot Do

Create, modify, or delete any record
Access fields outside the allow-list
Trigger actions, send messages, or initiate outreach
Access data from other organizations
Bypass rate limits or daily caps

Access Controls

Per-User Keys

Each staff member receives a unique access key stored in their Airtable account record. Keys are validated with timing-safe comparison on every request. There is no persistent session — revoking a key takes effect on the very next request.

Organization Isolation

Each MCP request resolves to a specific organization, which maps to a dedicated Airtable base. A user authenticated for one organization cannot query another organization's data under any circumstances. This is enforced at the credential routing layer.

Rate Limits and Extraction Prevention

Control	Limit
Per-minute rate limit	60 requests per key
Response size cap	100 records per response
Daily retrieval cap	5,000 records per key per 24-hour rolling window
Anomaly detection	High-frequency patterns flagged in audit trail

Normal conversational usage is typically 50-200 records per day. The daily cap prevents meaningful data extraction while supporting all legitimate usage patterns.

Data Minimization

The MCP server applies allow-list field filtering before returning data to the AI assistant:

Only explicitly approved engagement fields are included (donation amount, volunteer hours, event participation, communication history)
All other fields are blocked by default — including custom fields added after onboarding
Specific tools that need additional fields (e.g., export includes name and email) declare their allowed fields explicitly in code
This is deterministic and rule-based, not an AI judgment call

The AI assistant never sees full street addresses, phone numbers, dates of birth, or government identifiers — even if those fields exist in your Airtable base.

The LLM Data Path

When MCP returns data to the AI assistant, that data transits through the LLM provider:

Provider	Training on Your Data	Retention
Anthropic (Claude)	No — excluded by API terms	Up to 30 days for trust and safety
OpenAI (ChatGPT)	No — excluded by API policy	Up to 30 days for abuse monitoring

Zero-retention configurations are available depending on contract tier.

Audit Trail

Every MCP query is logged with: user key (hashed), timestamp, tool invoked, organization ID, records returned, and response time. Supporter PII is never logged.

Logs are available in Vercel function logs (1-hour retention) or via SIEM streaming (30-90 day retention) for organizations on high-compliance onboarding.

Disabling MCP

MCP is optional. To disable:

Remove individual access keys from Airtable account records
Or request that AI4Love disable MCP access for your entire organization

AI-generated insights continue appearing in the dashboard — they are written directly to Airtable by nightly agents and do not depend on MCP.

For full technical details, see the AI4Love Trust Center — MCP Access Model.

Common Questions

Does MCP mean AI has direct access to our database? No. The MCP server is an intermediary that validates access, filters fields, enforces rate limits, and logs every query. The AI assistant only sees what the MCP server explicitly returns after filtering.

Can staff see each other's queries? No. Each staff member has their own access key. Query logs are available to administrators but not to individual users.

What if we only want to use Claude, not ChatGPT? MCP access can be restricted to a single LLM provider.

Is there a way to audit what data the AI assistant has seen? Yes. Every MCP query is logged with the tool invoked, record count returned, and timestamp. SIEM streaming provides 30-90 day retention for forensic review.

Can the AI assistant send messages to our supporters? No. MCP tools are strictly read-only. The assistant can surface information and recommendations, but cannot take any action.