A Model Context Protocol server to help AI with Frappe app development
Frappe MCP demonstrates moderate definition quality with clear naming and structured schemas, but suffers from inconsistent parameter descriptions, missing output schema documentation, and inadequate error handling guidance. All 18 tools use proper verb_noun naming conventions (find_symbols, get_function_signature, create_doctype, run_bench_command). Schemas are well-defined with schemars JsonSchema derivations visible in source. However, parameter descriptions vary significantly in completeness, and no tool documents its output schema or return type structure. Tool descriptions are present but brief (10-50 chars typically), falling short of the 50-200 char production baseline. Risk classification exists (READ_ONLY, WRITE, IRREVERSIBLE) but error recovery guidance is absent. Per-tool review shows Rust struct definitions with proper serde attributes, but descriptions lack LLM-optimized context about when to use each tool versus similar ones.
DocType name to analyze relationships for
Page name (e.g., "User Settings", will be converted to snake_case)
DocType name (e.g., "Task")
Report name (e.g., "Sales Analysis")
DocType name (e.g., "Sales Invoice")
Slug prefix for the web, eg: "sales-portal", don't include "www/".
Symbol name to search for across the project/app source files
Output schemas not documented. No tool specifies what fields are returned or their types. Agents cannot plan downstream calls or extract relevant data.
Tool descriptions are too brief (10-50 chars). They lack context about WHEN to use each tool versus similar ones, and do not state WHAT the tool returns. Falling far short of 50-200 char production baseline.
Destructive/irreversible tools (run_bench_command, run_mariadb_command, run_bench_execute) lack confirmation or dry-run support. No guidance for error recovery or what to do if the operation partially fails.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 50 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 9 | - | v1 |
DocType name (e.g., "Sales Invoice")
DocType name (e.g., "Sales Invoice")
DocType name to search field usage in
Function name to find
Optional module filter to list DocTypes only from a specific module
Document ID (e.g., a7b9c3, d8f2e1) to read the full content of a Frappe documentation file
Arguments to pass to the `bench` command, eg: `migrate`, `mariadb -e "SELECT 1"`, etc.
Frappe function to execute, e.g., "frappe.db.get_list"
SQL query to execute via bench mariadb command
Specific module to test (optional)
Search query string
Error responses do not include recovery guidance. Stack traces and raw error codes give LLMs no actionable path forward. Missing error classification (retryable vs user-fixable vs fatal).
Parameter descriptions lack constraint details. E.g., search_in accepts 'backend', 'frontend', 'all' but the description does not formally list these as enum values. LLMs must infer valid options from text.
No pagination support visible in tools that return lists (find_symbols, list_doctypes, search_frappe_docs). No limit enforcement stated, risking context window exhaustion.
Security: run_bench_command and run_mariadb_command accept raw command/SQL strings with no validation. Vulnerable to command injection and SQL injection if LLM-generated input is not sanitized.
Composition: run_bench_execute accepts raw frappe_function strings (e.g., 'frappe.db.get_list'). This requires the agent to know Frappe API internals. Wrapping specific high-value functions (get_list, get_value, get_all) as dedicated tools would reduce agent complexity.