An MCP server providing SQL query execution, weather data, JavaScript code execution in sandboxed V8 isolates, and TypeScript type schema generation. Runs on Cloudflare Workers with Durable Objects storage.
This server has 6 tools with moderate quality. Tool names follow verb_noun convention well (sql_query, sql_exec, get_weather, execute_code). Descriptions exist but vary in quality, some are verbose and include implementation details rather than LLM-optimized guidance. Parameter schemas are present with types and basic descriptions, but lack constraints (enums, ranges, patterns). Output schemas are not documented. Error handling is partially present but lacks recovery guidance. The execute_code tool has an exceptionally long description (600+ chars) that reads like implementation docs rather than agent-facing guidance. SQL tools have minimal descriptions. No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels documented in the risk field.
Execute JavaScript in a sandboxed V8 isolate. The code has access to a `codemode` object with typed functions for all other tools. The code should be a function body (not wrapped in a function). Use `return` to return a value. For fast read-only SQL, use the built-in `query(sql, params)` helper. For multiple reads in one round-trip, use `queryBatch([{sql, params}, ...])`. To persist structured data into SQLite, use `store(tableName, arrayOfObjects)`.
Get TypeScript type definitions for all tools
Get current weather for any city. Returns temperature, humidity, wind speed, and weather condition code.
Execute DDL/DML (CREATE TABLE, INSERT, UPDATE, DELETE)
Execute multiple DDL/DML statements in one call
Execute SELECT queries against SQLite
Tool descriptions are not LLM-optimized. sql_query (35 chars), sql_exec (40 chars), and sql_exec_batch (50 chars) are too brief and lack action context or differentiation from similar tools. execute_code is 600+ chars and reads as implementation docs rather than agent guidance. Baseline is 50-200 chars.
No output schemas documented for any tool. Agents have no specification of response structure, field names, types, or nesting. This forces LLMs to infer structure from examples and makes chaining tools unreliable. Example: sql_query should document that it returns an array of objects with typed columns.
No tool annotations (toolAnnotations flag is false). Tools with clear risk levels (sql_exec, sql_exec_batch, execute_code are WRITE; sql_query, get_weather, get_type_schema are READ_ONLY) should declare readOnlyHint and destructiveHint in their definitions. This lets agents understand side effects without reading descriptions.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 53 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 41 | - | v1 |
Parameter constraints are missing. sql_query and sql_exec accept free-form SQL strings with no validation guidance. Agents may pass UPDATE/DELETE to sql_query or SELECT to sql_exec. Add validation notes to descriptions: 'Must be a valid SELECT statement (INSERT/UPDATE/DELETE not allowed)' and provide error messages on constraint violation.
Tool descriptions do not differentiate between similar tools. sql_query, sql_exec, and sql_exec_batch all operate on the same database. An agent does not know from descriptions alone when to pick one vs the others. Add dependency guidance: 'Use sql_query for SELECT. Use sql_exec for single DDL/DML (CREATE TABLE, INSERT, UPDATE, DELETE). Use sql_exec_batch for multiple statements in one call to reduce latency.'
execute_code parameter description is implementation-focused rather than agent-focused. It includes code snippets, helper function names, and examples that belong in separate documentation, not the parameter schema. Descriptions should explain WHAT the parameter does, not HOW to use internal APIs. Refactor: 'JavaScript code (as function body) to execute in a V8 isolate. Can access codemode object for tool calls, query/queryBatch helpers, and store function.'
No error handling documentation. Tools do not specify what errors can occur, whether they are retryable, or what the LLM should do next. Example: sql_query might fail with 'Table not found' or 'Syntax error', agent needs recovery guidance for each.