An MCP server that provides tools for interacting with Salesforce via SOQL queries, Tooling API, metadata retrieval, and object description
This Salesforce MCP server has clear, well-structured tool definitions with explicit schemas and input validation logic. However, it suffers from critical deficiencies in parameter documentation, output schema specification, and error handling guidance that would impede LLM usage in production. All four tools are explicitly registered in src/tools.ts with inputSchema objects and handler functions, making definitions directly visible (not inferred). Tool names are action-oriented (query, describe-object, metadata-retrieve) but lack the prefix conventions (get_, list_, create_) that help LLMs parse intent. Most critically, ZERO parameter descriptions exist for the main inputs, 'query' and 'objectName' parameters have descriptions, but 'type', 'fullNames', and 'detailed' lack context beyond the JSON Schema. Output schemas are completely undocumented; handlers return raw jsforce results without specifying what fields agents should expect. Error messages are generic (e.g., 'Invalid describe object arguments') with no recovery guidance. The server does validate input types (isValidToolingQueryArgs, isValidMetadataType) but doesn't leverage that validation for user-facing error feedback.
Get detailed metadata about a Salesforce object
Retrieve metadata components from Salesforce
Execute a SOQL query on Salesforce
Execute a query against the Salesforce Tooling API
Output schemas are completely undocumented. Tools return raw jsforce Connection results (e.g., conn.query(), conn.metadata.read()) without specifying what fields, types, or structure agents should expect. This forces LLMs to infer response format and wastes tokens.
Parameter descriptions are sparse or missing context. 'fullNames' (metadata-retrieve) has only 'Array of component names to retrieve', no guidance on format, escaping, or limits. 'type' parameter lacks explanation of when to use each metadata type. 'detailed' flag's impact is vague ('Whether to return full metadata'), does it add latency? What extra fields appear?
Tool names lack action verb prefixes. 'query' and 'tooling-query' omit 'get_' or 'search_' prefix; 'describe-object' uses dashes instead of underscores (inconsistent with MCP convention verb_noun). LLMs scanning a tool list struggle to recognize what these tools do vs. similar tools.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 51 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 38 | - | v1 |
Error handling provides no recovery guidance. Generic errors like 'Invalid metadata retrieve arguments' or 'Invalid metadata type: ...' don't tell the LLM what to do next. No mention of retryability, available alternatives, or required preconditions.
No pagination support or result limits. 'query' and 'tooling-query' accept free-form SOQL without limit hints. Large result sets risk context window exhaustion. Handlers return entire result objects without capping records or offering cursor-based pagination.
Salesforce credentials (username, password, security token) are required environment variables with no mention of how they're validated or rotated. While not exposed as tool parameters (correct), the server logs '[MCP Error]' to stderr without sanitizing sensitive context from error messages.