Atlan

Exposes search and discovery tools via the Model Context Protocol that translate MCP tool calls into pyatlan SDK queries against the Atlan metadata platform. Uses a FastMCP server core that routes structured search requests through access-control middleware before dispatching to asset discovery modules, enabling AI agents to query data lineage, ownership, classifications, and custom metadata fields without direct API knowledge.

Implements a lineage traversal tool that accepts an asset identifier and traverses upstream (source) and downstream (dependent) data flows through the Atlan metadata graph. Uses pyatlan SDK to fetch lineage relationships and exposes them as structured tool outputs, allowing AI agents to understand data provenance, impact analysis, and transformation chains without manual graph database queries.

Implements the MCP server core using the FastMCP framework, which provides a decorator-based tool registration system (@mcp.tool()) and automatic MCP protocol handling. Tools are registered as Python functions with type-annotated parameters, and FastMCP automatically generates MCP tool schemas, handles protocol serialization, and routes incoming tool calls to implementations. The server instantiates FastMCP, registers 15 tools across discovery, lineage, update, glossary, quality, and domain domains, and selects transport mode at startup.

Provides a Docker image (ghcr.io/atlanhq/atlan-mcp-server) that packages the MCP server with all dependencies, enabling single-command deployment without local Python setup. The image includes the atlan-mcp-server package, pyatlan SDK, FastMCP, and all dependencies, and accepts configuration via environment variables passed at container runtime. Supports multiple transport modes (stdio, HTTP) and can be deployed to Kubernetes, Docker Compose, or cloud container services.

Distributes the Atlan MCP server as a Python package (atlan-mcp-server) on PyPI, enabling installation via pip without cloning the repository. Package includes all source code, dependencies, and entry points for running the server locally or in development environments. Supports installation with pip install atlan-mcp-server, making it accessible to Python developers and enabling integration into existing Python projects.

Implements helper functions (parse_json_parameter(), parse_list_parameter()) that parse string-based tool inputs into structured Python objects. Handles JSON deserialization for complex parameters and list parsing for comma-separated or JSON array inputs, enabling MCP clients to pass structured data as strings and tools to receive typed Python objects. Provides error handling for malformed JSON and invalid list formats.

Provides an asset update tool that accepts asset identifiers and metadata patches (key-value pairs for custom attributes, descriptions, owners, classifications) and applies them via the pyatlan SDK's batch update mechanism. Validates input schemas against Atlan's asset type definitions before submission, preventing malformed updates and providing structured error feedback to the agent.

Exposes glossary management tools that enable AI agents to create, read, update, and delete business glossary terms within Atlan's hierarchical glossary structure. Tools support term creation with parent-child relationships, attribute assignment, and linking terms to data assets, allowing agents to build and maintain business metadata catalogs programmatically through MCP protocol calls.

Exposes data quality tools that allow AI agents to define quality rules (e.g., null checks, uniqueness constraints, value range validations) and link them to assets, then query rule execution results and anomaly detection outcomes. Rules are stored in Atlan's quality module and executed by connected data systems, with results aggregated and exposed via MCP tools for agent-driven quality monitoring and alerting.

Exposes tools for managing Atlan's domain and product hierarchies, allowing AI agents to create domains (logical data groupings), assign assets to domains, and organize products within domains. Domains provide hierarchical asset organization and access control boundaries, enabling agents to structure metadata catalogs by business domain or data product without manual UI interaction.

Implements a ToolRestrictionMiddleware layer that intercepts MCP tool calls and filters tool visibility and execution based on environment-configured RESTRICTED_TOOLS list. Tools listed in restrictions are either hidden from the MCP client's tool list or blocked at execution time, enabling fine-grained access control without modifying tool implementations. Configuration is loaded from environment variables or .env files, allowing deployment-time tool restriction without code changes.

Supports multiple transport modes (stdio, HTTP/SSE, WebSocket) for MCP client connectivity, allowing the same FastMCP server to be deployed locally (stdio for Claude Desktop, Cursor IDE) or as a remote service (HTTP for web-based agents). Transport selection is configured via environment variables, enabling single codebase deployment across different client architectures without code changes.

Implements a singleton factory pattern (get_atlan_client()) that creates and caches a single pyatlan AtlanClient instance across all MCP tool calls, managing authentication credentials loaded from environment variables (ATLAN_API_TOKEN or ATLAN_USERNAME/ATLAN_PASSWORD). Reuses the same authenticated client across tool invocations, reducing authentication overhead and connection pooling overhead while centralizing credential management.

Implements a Pydantic BaseSettings-based configuration system (Settings class, get_settings() singleton) that loads all MCP server configuration from environment variables or .env files. Supports configuration for Atlan credentials, restricted tools, transport mode, and server port, with type validation and default values. Configuration is loaded once at startup and cached, enabling flexible deployment without code changes.