Add a connection. Map your tables. Ask a question.
Create an account at pathbound.ai. Every plan starts with 14 days free, no card.
In the dashboard, go to Integrations → PostgreSQL and enter the host, port, database, user, and password of a read-only user. TLS is supported, with your own CA certificate, and the server has to accept connections from outside your network. Pathbound reads your schema: for each table you want, choose what a row is and how it matches—by email, domain, or id—then select Sync now.
Drop the snippet below into your Claude Desktop config (or in Claude.ai web: Customize → Connectors → Add custom connector → enter https://mcp.pathbound.ai/mcp). The same URL works for ChatGPT, Cursor, Cline, and any MCP client that supports OAuth.
{
"mcpServers": {
"pathbound": {
"url": "https://mcp.pathbound.ai/mcp"
}
}
}Ask the agent: “Which trial accounts have the most active seats, and what did they last ask support?” The agent reads the synced fields beside each account’s CRM and support history.
Plans, seats, usage, and invoices live in your own database. Pathbound reads the tables you choose, resolves each row to the right contact or company, and serves it beside everything else over the Model Context Protocol—so an agent asked about an account sees what it pays for and how it uses the product.
Pathbound discovers your schema—tables, columns, keys, and foreign keys. You decide what a row is (a contact, a company, an event, or a related record such as a subscription) and how it matches. No ETL job to write or maintain.
Rows go through the same identity resolution as every other source, so a user in your database, a contact in your CRM, and a requester in your help desk become one profile.
Pathbound runs only SELECT queries it builds itself, each with a time limit, and never writes to your database. Connect it with a read-only database user.
Mapped columns become profile fields agents can filter, count, and segment on—“trial accounts with more than ten seats”—and the full row stays on the profile for context.
No. Pathbound runs only SELECT queries it builds itself, each with a time limit, and never writes. We still recommend a database user that can only read, so the database enforces it too.
No. Agents never query your database. They read what Pathbound synced, through the same tools as every other source: profiles, filtered lists, aggregates, and segments.
Any table the database user can read. A table can become contacts, companies, events, or records attached to a contact or company, such as subscriptions or invoices. Columns from a related table can be folded onto the same row through a foreign key.
On a schedule, and at once when you select Sync now in the dashboard. Each sync updates the mapped fields on the profile, and the full row stays available to agents beside them.
Tenant-isolated, with your database password encrypted at the application layer and TLS to your server. Pathbound does not sell customer data, and we don’t call AI providers on your behalf—your MCP client does, under your account and the AI provider’s terms.
Same MCP server, same tools—just connect a different source. See the integrations directory.
Sign up, connect one CRM or drop the events snippet, and point your MCP client at mcp.pathbound.ai/mcp.