AWS Nova Canvas

Exposes Amazon Nova Canvas image generation as an MCP tool that LLM clients can invoke through standardized tool-calling interfaces. The server implements the MCP protocol's tool schema registration pattern, allowing Claude, other LLM clients, and AI agents to call image generation with structured input validation and streaming response handling. Requests are translated to AWS Bedrock InvokeModel API calls with proper credential management via AWS SDK.

Accepts natural language prompts and optional structured color guidance parameters to control the visual output of generated images. The server parses color specifications (hex codes, RGB values, or semantic color names) and passes them to Nova Canvas as generation parameters, enabling fine-grained control over image aesthetics without requiring multiple generation attempts.

Implements the MCP server protocol lifecycle including initialization, tool schema registration, request routing, and graceful shutdown. The server registers image generation as a callable tool with JSON schema validation, handles incoming MCP requests through stdio transport, and manages the connection state with MCP clients. Uses Python's asyncio for concurrent request handling and proper error propagation back to clients.

Abstracts AWS Bedrock InvokeModel API calls through boto3 SDK, leveraging AWS credential chain (IAM roles, environment variables, or credential files) for transparent authentication. The server constructs Bedrock API payloads with Nova Canvas model parameters, handles streaming responses, and translates Bedrock errors into MCP-compatible error responses. Supports both synchronous and asynchronous invocation patterns.

Encodes generated images as base64 strings for transmission through MCP protocol (which uses JSON-RPC 2.0 over stdio). The server handles image data from Bedrock, applies base64 encoding, and embeds the encoded data in MCP response payloads along with metadata (dimensions, generation parameters). Supports both inline embedding and optional S3 URL references for large images.

Defines and enforces JSON schema for image generation tool inputs, validating prompt text, color parameters, and optional generation settings before passing to Bedrock. The server uses schema validation to reject malformed requests early, provide meaningful error messages to clients, and ensure type safety. Schema is registered with MCP tool definition and enforced at request time.

Catches Bedrock API errors (throttling, authentication failures, model unavailability) and translates them into MCP-compatible error responses with descriptive messages. The server implements exponential backoff for transient errors, distinguishes between client errors (invalid parameters) and server errors (service unavailable), and propagates error context to help debugging. Errors are formatted as MCP error objects with error codes and messages.

Uses Python asyncio to handle multiple concurrent image generation requests without blocking. The server implements async/await patterns for Bedrock API calls, allowing multiple clients to submit requests simultaneously and receive responses independently. Concurrent requests are queued and processed based on system resources and Bedrock API rate limits, enabling efficient resource utilization.

Registers image generation as a discoverable MCP tool with complete JSON schema definition, enabling MCP clients (Claude, other LLMs) to introspect available capabilities and understand expected input/output formats. The server exposes tool metadata including name, description, input schema, and usage examples, allowing clients to make informed decisions about when and how to invoke image generation. Tool schema follows MCP specification for standardized client integration.