An MCP server that exposes GraphQL operations as tools, enabling GraphQL queries and mutations to be executed as MCP tools with introspection and generic operation execution capabilities
This GraphQL MCP server exposes three tools with significant quality gaps. The introspect and operation tools have extremely minimal descriptions (under 50 chars each), making them difficult for LLMs to reason about when to use them. Input schemas are present but underdescribed, the operation tool accepts a raw 'operation' string with no guidance on valid GraphQL syntax, format constraints, or examples. The dynamic_graphql_operations tool is problematic: it's described as a single tool, but actually represents N tools generated from config files. The generated tools' definitions are inferred rather than explicitly visible in code, preventing verification of their individual schemas and descriptions. Parameter descriptions are largely absent (only 'operation' has a description in operation tool). No output schemas are documented anywhere. Error handling is not visible in the provided code. The server lacks tool-level annotations for distinguishing read-only vs write operations beyond the Risk field (which is metadata, not schema). Overall, this server exhibits patterns typical of early-stage integrations: minimal self-documentation, no input validation guidance, no output shape documentation, and tool composition that conflates multiple operations under one label.
Dynamically generated tools from GraphQL operations defined in configuration files. Each operation becomes a tool with parameters based on its GraphQL variables.
Introspect the Evolve graphql schema
Perform an operation on the Evolve graphql schema. Can be either a query or a mutation.
Tool descriptions are far too short (10-30 chars). 'Introspect the Evolve graphql schema' and 'Perform an operation on the Evolve graphql schema' provide almost no context for LLM selection. No explanation of when to call introspect vs operation, or what 'Evolve' is.
dynamic_graphql_operations conflates multiple tools under one definition. The description claims 'Each operation becomes a tool' but the tool itself is singular. Individual generated tools' schemas, descriptions, and parameter types are not visible in source code. This violates the single-responsibility principle and prevents verification of tool quality.
The operation tool accepts a raw GraphQL operation string with no input validation guidance. No description of valid syntax, format constraints, mutation vs query distinction, variable handling, or error modes. LLMs will guess at valid input, leading to malformed queries.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 37 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 29 | - | v1 |
No output schemas documented for any tool. What does introspect return? Is it a schema definition, a list of types, or raw introspection JSON? What does operation return, query result, mutation confirmation, error? LLMs cannot plan multi-step sequences without knowing response shape.
Parameters lack descriptions. The operation tool's 'operation' parameter has a description, but generated tools from dynamic_graphql_operations have no visible parameter documentation. Variables inferred from GraphQL schema are not self-documenting to LLMs without descriptions.
No error handling guidance visible. How does the tool respond when a GraphQL query is malformed? When the endpoint is unreachable? When a mutation fails? LLMs receive no recovery path, no distinction between retryable vs fatal errors.
Tool naming violates verb_noun convention. 'introspect' and 'operation' are vague. 'introspect' should be 'get_schema_introspection' or 'describe_graphql_schema'. 'operation' should be 'execute_graphql_query', 'execute_graphql_mutation', or split into separate tools.
The operation tool combines both queries and mutations under one interface. LLM cannot infer from the tool name whether it's safe to call (read-only) or destructive. Should be split into execute_graphql_query (read-only) and execute_graphql_mutation (write) with appropriate tool annotations.
Tool annotations are present (WithReadOnlyHintAnnotation, WithOpenWorldHintAnnotation applied to generated tools in additional.go) but NOT applied to the built-in introspect and operation tools. introspect should have readOnlyHint=true; operation should have different hints depending on query vs mutation type, but this is not visible in the tool definition.