A mock HTTP server implementing the Agentic Recruitment Protocol (ARP) specification, providing REST and JSON-RPC 2.0 endpoints for job posting, application, offer, and interview management
This is a mock MCP server for the Agentic Recruitment Protocol. Tool definitions are present with schemas and descriptions, but quality is inconsistent. All 8 tools have names starting with action verbs (search, create, get, book, update, grant, revoke), which is positive. Descriptions range from 50-130 characters, adequate but sparse for LLM disambiguation. Input schemas are visible in the code and properly typed (string, integer, array, object), meeting the minimum. However, parameter descriptions are generic or missing context about constraints, dependencies, and error conditions. Output schemas are partially documented (return structures visible in code but not in formal schema declarations). Error handling is present but minimal, most errors are HTTP status codes without actionable recovery guidance. The tool set is domain-appropriate for recruitment but lacks composition clarity around candidate lookup, consent tracking, and interview scheduling workflows.
Book an interview slot for an application
Create a new job application for a candidate to a job posting
Create a job offer for a candidate application
Retrieve details of a specific application by ID
Grant consent for data processing on behalf of a candidate
Revoke previously granted consent scopes for a candidate
Search and filter job postings by query, location, employment type with pagination support
Update an existing offer with counter offer, acceptance, or decline actions
Parameter descriptions lack constraint specifications. E.g., 'slot' in bookInterview is described as 'Interview slot datetime in ISO format, must be in the future' but lacks explicit format pattern or validation rule. 'timezone' defaults to UTC but the impact on scheduling logic is unclear. LLMs cannot infer timezone handling rules from natural language alone.
No output schemas documented. While code shows responses (e.g., createApplication returns {id, state, status_url}), there is no formal JSON Schema declaration of return types. LLMs cannot infer which fields are guaranteed, optional, or conditional without explicit schema.
Error handling is generic HTTP status codes (400, 404, 500) without structured error responses or recovery hints. Example: getApplication returns 404 with {error: 'Not found'}, no guidance on how to find the correct application ID or whether to retry.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 48 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 35 | - | v1 |
Descriptions are under 100 characters for 6 of 8 tools (getApplication: 50 chars, grantConsent: 55 chars, revokeConsent: 50 chars). Baseline for production tools is 194 characters. These are too terse to disambiguate from similar consent/retrieval operations. Example: getApplication ('Retrieve details of a specific application by ID') does not explain when to call this vs searchJobs or createApplication.
Parameter 'terms' in createOffer is typed as 'object' with description 'Offer terms including salary, benefits, etc. (required)', no schema for the object shape. Example values ('currency', 'base') are in code but not in parameter docs. LLMs cannot construct valid terms objects without explicit field definitions.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). searchJobs and getApplication are read-only; createApplication, bookInterview, createOffer, updateOffer, grantConsent are write operations; revokeConsent is reversible. These annotations are missing from tool definitions, preventing agents from reasoning about side effects.
No idempotency guarantees declared. bookInterview and createOffer could generate duplicate records if retried due to transient errors. No mention of idempotent keys or deduplication logic. Agents cannot safely retry without risk of side effects.
searchJobs returns paginated results but limit and offset logic lacks bounds checking in documentation. Code enforces max limit of 100, but this constraint is not stated in the parameter description. LLMs may pass unreasonable values expecting them to work.
employmentType parameter in searchJobs is an array but no enum values or examples documented. Code checks for 'FULL_TIME' but parameter description says 'e.g., FULL_TIME', no list of valid types (FULL_TIME, PART_TIME, CONTRACT, etc.).
updateOffer 'action' parameter accepts COUNTER, ACCEPT, DECLINE but no enum constraint documented. Free-form string invites invalid values. Should declare enum: ['COUNTER', 'ACCEPT', 'DECLINE'].