chaining-mcp-server

Enables MCP clients to execute multiple tools in sequence where output from one tool automatically becomes input context for the next tool in the chain. Implements a state machine pattern that tracks execution flow, maintains intermediate results, and handles error propagation across chain steps. The server parses tool invocation requests, validates schemas at each step, and ensures context is properly serialized between tool boundaries.

Validates incoming tool invocation requests against registered tool schemas and transforms parameters between chain steps using schema-aware mapping. The server maintains a registry of tool definitions with JSON Schema validation, performs type coercion where safe, and rejects malformed requests with detailed error messages indicating which schema constraint was violated. Supports parameter renaming and structural transformation between tool boundaries.

Manages execution context and intermediate state across tool invocations within a single MCP session, storing results from each tool step and making them available to subsequent steps. The server maintains an in-memory context store keyed by execution session ID, serializes intermediate results, and provides context retrieval mechanisms for tools that need access to prior outputs. Supports context scoping (global vs. chain-local) and automatic cleanup of stale context.

Detects tool execution failures, captures error context, and provides configurable recovery strategies (fail-fast, retry with backoff, skip-and-continue, fallback-tool). The server implements error classification (schema validation errors vs. tool execution errors vs. timeout errors), logs detailed error traces with execution context, and allows chains to define per-step error handlers. Supports conditional error recovery based on error type and retry count.

Maintains a registry of available tools with metadata (name, description, schema, version) and exposes tool discovery mechanisms to MCP clients. The server implements tool registration endpoints, supports tool versioning, and provides introspection APIs that allow clients to query available tools, their schemas, and capabilities. Tools can be registered at server startup or dynamically added at runtime through registration endpoints.

Records detailed execution traces for each tool chain invocation, including tool names, parameters, results, timing, and error information. The server implements structured logging with execution IDs, captures intermediate state at each step, and provides trace retrieval APIs that allow clients to inspect execution history. Traces include timing information for performance analysis and can be exported in standard formats (JSON, structured logs).

Enables execution of independent tools in parallel within a single chain, with automatic dependency resolution to ensure tools execute in the correct order. The server analyzes tool parameter dependencies, identifies tools that can run concurrently, and executes them in parallel using async/await patterns. Supports explicit dependency declarations and automatic dependency inference from parameter references.

Supports conditional execution paths within tool chains based on tool outputs, allowing chains to branch based on results and route to different tools dynamically. The server evaluates conditions (equality, comparison, regex matching) against tool outputs, selects the appropriate next step, and continues execution on the selected branch. Supports nested conditions and fallback paths when no condition matches.