Add Model Context Protocol Support to Plumber APIs. Enables AI assistants to discover and interact with R functions through MCP tools, resources, and prompts.
plumber2mcp exposes 4 tools derived from Plumber API endpoints. All tools have descriptions and complete input schemas with proper typing. However, several quality issues significantly limit production readiness: (1) Tool names violate verb_noun conventions, GET_echo, POST_add, GET_time, POST_parse use HTTP method prefixes instead of action verbs (should be 'echo', 'add', 'get_time', 'parse_json'). This breaks LLM naming heuristics. (2) Descriptions are extremely terse (13-21 chars), below the 50-200 char target for LLM optimization. No descriptions explain WHEN to use each tool or WHAT happens after calling it. (3) Parameter descriptions are minimal: 'First number', 'Second number' provide no context about valid ranges, units, or constraints. (4) Output schemas are not documented, LLMs cannot predict what fields they receive or plan downstream tool chains. (5) No error guidance, tools are stateless HTTP forwarders with no recovery hints if parsing or computation fails. (6) Default parameter values (empty string for 'msg', omitted 'a' and 'b') invite silent failures. These are fixable but require material revision to meet production baselines.
Echo back the input
Get current time
Add two numbers
Parse JSON data
Tool names violate verb_noun convention and use HTTP method prefixes (GET_, POST_) instead of action verbs. LLMs rely on names to infer intent before reading descriptions. 'GET_echo' should be 'echo', 'POST_add' should be 'add', 'GET_time' should be 'get_time', 'POST_parse' should be 'parse_json'. This forces LLMs to read full descriptions for every tool, wasting tokens and increasing selection errors.
Tool descriptions are 13-21 characters, far below the 50-200 char baseline for LLM-optimized descriptions. Examples: 'Echo back the input' (19 chars), 'Add two numbers' (15 chars), 'Get current time' (16 chars), 'Parse JSON data' (15 chars). These lack context about use cases, prerequisites, side effects, and expected output structure. LLMs cannot distinguish when to select each tool without reading the schema.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 32 | 2025-06-18+ | v2 |
| 2026-03-09 | F | 16 | - | v1 |
Output schemas are not documented in tool definitions. LLMs have no way to predict what fields the tools return or plan multi-step chains. Per the rubric, 100% of A+ tools have documented return types. Tools must document: return type (string, object, array), field names, field types, and field descriptions. Without this, agents cannot reliably compose tools.
Parameter descriptions are single words or short phrases with no constraint information. 'msg' param: description is 'The message to echo' but provides no max length, encoding, or example format. 'a' and 'b' params: descriptions are 'First number' and 'Second number' but omit valid range, precision (int vs float), or units. Per the rubric, descriptions must state format, range, and constraints inline.
GET_echo has optional parameter 'msg' with default empty string. This silently invokes the tool with no user input, likely returning an empty echo. No error or warning guides the LLM to provide a message. Similarly, POST_add lists 'a' and 'b' as required but omits guidance if parsing fails. Default empty strings invite silent failures.
No error handling or recovery guidance. If POST_parse receives invalid JSON, the tool returns an HTTP error with no hints for the LLM to retry, ask the user for valid JSON, or call a different tool. Per the rubric, error responses must tell the LLM what to do next. This is a silent failure path.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). All tools are marked risk: READ_ONLY, but this metadata is not surfaced in the tool schema or MCP capability declaration. LLMs cannot infer idempotency or safety constraints from names alone.