Settra

Make tabular file data durable and easy for automated agents to use.

Settra is a self-hosted MCP server that synchronizes source data into PostgreSQL and exposes durable snapshots through a governed semantic layer. Automated agents can discover exact schemas, inspect bounded samples, and run structured queries without raw SQL or direct access to source credentials.

It is built for teams that want agents to work with operational data safely, consistently, and repeatably. The current release supports Google Sheets, CSV, Excel, and Parquet files selected from Google Drive.

Sources and destinations are modeled separately. Each pipe connects one Drive file to a registered destination and target namespace. Today Settra seeds one explicit Managed PostgreSQL destination backed by the deployment's POSTGRES_* settings, leaving a clean boundary for additional destination types later.

What can agents do?

• Find overdue items in an operations tracker.

• Summarize this month's pipeline from a sales worksheet.

• Compare actual values with targets stored in another tab.

• Identify rows that changed or need follow-up.

• Reuse an approved definition such as “active customer” or “recognized revenue” in later queries.

Related MCP server: mcp-data-lens

How it works

flowchart LR
    sheet["Google Drive tabular files<br/>Sheets, CSV, Excel, Parquet"]
    sync["dlt full sync<br/>OAuth + loading rules"]
    postgres["PostgreSQL<br/>Durable snapshots"]
    settra["Cube + Settra<br/>Approved semantics"]
    agent["Automated agent<br/>MCP client"]
    task["Question or workflow"]

    task --> agent
    agent -->|"structured metadata and queries"| settra
    sheet --> sync
    sync --> postgres
    postgres --> settra
    settra -->|"bounded results"| agent

Settra detects the file format and initial CSV delimiter, encoding, and header row, then performs a complete replacement load with dlt. PostgreSQL keeps the last successful snapshot available while a new one is staged. Per-source YAML controls parsing overrides, selected sheets, schedules, types, names, schema contracts, and descriptions.

The canonical semantic layer gives agents stable names, measures, dimensions, business definitions, and validation rules instead of unrestricted SQL access.

Collections group related pipes into focused agent workspaces. An agent using the global MCP URL asks which collection to use, loads its context once, and queries only its derived destination tables and cubes. A collection-specific MCP URL can optionally pin that selection.

Each account starts with a private personal workspace. Connections, collections, Google authorization, semantic assets, Cube queries, MCP grants, and request metrics are isolated to that workspace. The membership model is ready for shared organization workspaces without changing object ownership.

How data is handled

When self-hosted, Settra runs inside infrastructure you control. The Google OAuth refresh token is encrypted per workspace with SECRET_KEY on the data volume and is not stored in the product database or source YAML. MCP request/response contents are also not stored; PostgreSQL request history contains privacy-safe usage metrics only.

Query results are sent to the AI assistant or agent you connect, so that provider's privacy and retention policies still apply.

What you need

• A Google Sheet, CSV, Excel, or Parquet file in Google Drive.

• A Google Cloud Web OAuth client plus a browser-restricted Google Picker API key. Settra requests file-specific access only to files users select.

• A Settra deployment.

• An MCP-compatible AI assistant or automated agent.

For developers

• Self-hosting and technical setup

• Product database and migrations

• Google Cloud OAuth and Picker setup

• Architecture and API reference

• Google Drive source setup guide

• Contributing

Settra is open source under the Apache License 2.0.