The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The documentation provides instructions for users to download and install Materialize command-line tools and AI agent skills from official vendor domains and the associated GitHub repository.
Evidence: Instructions use curl to fetch binaries from binaries.materialize.com and npx to install skills from the MaterializeInc/agent-skills repository.
[INDIRECT_PROMPT_INJECTION]: The skill describes functionality for connecting AI agents to Materialize through MCP servers, allowing them to execute query and read_data_product tools. This creates an attack surface where an agent could ingest untrusted data from a user's database tables or Kafka topics, potentially leading to the execution of malicious instructions embedded in that data.
Ingestion points: Database tables, views, and Kafka topics accessed via MCP tools described in developer-tools/mcp-server/mcp-agent-tools/index.md.
Boundary markers: The documentation recommends setting restrict_to_user_objects = true on agent roles to prevent them from accessing sensitive system catalog metadata.
Capability inventory: The tools allow for arbitrary SQL SELECT statement execution and data retrieval within the assigned cluster.
Sanitization: Relies on external database-level RBAC and LLM guardrails; the skill provides detailed guides for setting up least-privilege roles.
[CREDENTIALS_UNSAFE]: Static analysis flagged a database connection string with embedded credentials in a guide for building custom sinks. These are typical in documentation as placeholders or local development examples (e.g., postgres://user:password@localhost), but they represent a minor security risk if users adopt them without modification.
Evidence: Detected in export-data/build-your-own-sink/postgres/index.md.
[NO_CODE]: The skill consists entirely of documentation in Markdown format and does not contain any executable scripts or binary files to be run directly by the agent platform.

mz-docs