Aegis Dq

What aegis-dq/aegis-dq MCP server does

The aegis-dq/aegis-dq MCP server exposes Aegis DQ’s data-quality pipeline through MCP. An agent can load a pipeline manifest and start validation without repeatedly supplying the rules, database location, knowledge-base files, or business goal. The pipeline can validate warehouse data against structured rules and return a report that distinguishes passed and failed checks.

Aegis DQ supports rules for completeness, uniqueness, accepted values, referential integrity, statistical checks, custom SQL, and machine-learning anomaly detection. Its documented warehouse adapters cover DuckDB, PostgreSQL and Redshift, BigQuery, Databricks, AWS Athena, and Snowflake. The MCP layer is intended for conversational use with Hermes, Claude Desktop, and other MCP-compatible agents.

How it works

Create a pipeline manifest that points to a rules file and database, then optionally add policy or schema documents as knowledge-base context. A goal can describe the explanation or remediation detail expected from the run. The agent first calls load_pipeline, followed by run_validation.

During validation, Aegis DQ evaluates the configured rules and can send failures to an LLM for diagnosis and root-cause analysis. Supported LLM options documented by the project include Anthropic Claude, OpenAI, Ollama, and AWS Bedrock. The workflow can also produce proposed SQL fixes; these are proposals in the documented pipeline and should be reviewed before execution.

The project records LLM decisions in an audit trail, including cost and latency. For environments that do not need model analysis, Aegis DQ supports offline validation with LLM processing disabled.

Setup and configuration

Install the Python package with:

pip install aegis-dq

Install the mcp extra for MCP support:

pip install 'aegis-dq[mcp]'

Start the MCP process with aegis mcp. A Hermes configuration example passes ANTHROPIC_API_KEY through the server environment. The required credential depends on the selected LLM provider; offline mode does not require an API key. Warehouse-specific extras are available for BigQuery, Databricks, Athena, PostgreSQL or Redshift, and Snowflake.

A minimal manifest can specify a name, rules path, database path, knowledge-base files, and a natural-language goal. The README demonstrates a local DuckDB file, but the framework also documents adapters for external warehouses.

Tools and capabilities

The documented MCP workflow provides:

• load_pipeline to load a pipeline manifest.
• run_validation to execute the loaded pipeline.
• Rule evaluation across supported warehouse adapters.
• LLM-based failure diagnosis and root-cause analysis when enabled.
• Generated SQL remediation proposals.
• Structured validation reporting.
• Audit records for LLM decisions, with cost and latency information.
• Offline validation when the LLM is disabled.

The same package can also be used outside MCP through its CLI, GitHub Actions integration, Airflow support, and REST extra, but those are separate interfaces from the MCP server.

Limitations and notes

The aegis-dq/aegis-dq MCP server does not remove the need to configure a pipeline, rules, and a warehouse connection. LLM diagnosis requires a supported provider and its credentials unless offline mode is selected. LLM-generated remediation SQL should be reviewed before applying it to production data.

The README identifies Claude Desktop as an MCP-compatible client, but it does not provide a complete standalone Claude Desktop configuration in the supplied material. Warehouse availability can also depend on installing the matching optional package extra and supplying the connection details required by that warehouse.

Aegis DQ is released under Apache-2.0. The project requires Python 3.11 or newer according to its package information.