An MCP server exposing the full read/write surface onto one hh.ru account (search, resumes, applying, chat, archiving) over plain HTTP. No official API, no browser, no token — uses session-based HTTP client to interact with hh.ru website.
Six read-only tools with inconsistent quality. Tool names lack action verbs (sent, lost, inbox, chats, whoami, resume_scorecard are nouns or state descriptors, not verb_noun patterns). Descriptions are verbose (200-400+ chars) and domain-specific but lack LLM-optimized structure. Input schemas present for most tools but sparse (1-2 params each). No output schemas documented. Error handling is minimal, tools return free-text strings rather than structured objects. No parameter type validation visible. Descriptions bury key details in implementation notes rather than leading with WHAT/WHEN/WHY.
Every messenger chat, including employer-initiated ones with no application. `sent`/`inbox` are built on /applicant/negotiations, which only knows chats of type NEGOTIATION. Employers writing to you first create type COMMON (subType GENAI = hh's AI recruiter), which those views cannot see at all.
Everything awaiting a reply. One cheap list; read a single chat for detail. Combines negotiations for state and chats for who-spoke-last, joined on chat_id.
Negotiations hh dropped from `sent`/negotiations() entirely, usually a rejection. hh eventually drops a topic that resolves to DISCARD -- verified 2026-08-30, absent from every page even though the chat and the rejection text survive in the messenger backend.
hh's own verdict on every CV, next to how each one actually performs. Pulls together three things hh keeps in separate places: the funnel (impressions / opens / invitations), `fieldStatuses.leftToFillFields` (hh's literal checklist of what is still empty), and the canonical-vs-freetext split of the skills.
Every application with its real state, across all pages.
Tool names are nouns/state descriptors, not action verbs. 'sent', 'lost', 'inbox', 'chats', 'whoami', 'resume_scorecard' lack verb_noun structure (e.g., list_sent_applications, get_lost_negotiations, list_inbox_items). LLMs cannot infer intent from names alone.
No output schemas documented. Tools return free-text strings (e.g., 'nothing matches', formatted table rows) instead of structured JSON objects. LLMs cannot parse results reliably or chain to downstream tools.
Three tools (lost, whoami, resume_scorecard) have empty input schemas ({}). No parameters documented. Descriptions do not explain what inputs are accepted or required.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 39 | 2026-07-28+ | v2 |
Which hh account this session is authenticated as. Call it first when anything looks wrong: the usual cause is a stale or borrowed session.
Descriptions are verbose (200-400+ chars) and implementation-focused, burying key details. 'lost' description is 400+ chars of caveats about hh.ru's behavior; 'inbox' is 600+ chars of performance notes. LLM-optimized descriptions should be 50-200 chars, leading with WHAT/WHEN.
No error handling guidance. Tools return bare strings ('nothing matches', 'nothing awaiting a reply') without recovery hints. LLMs cannot distinguish retryable failures from empty results.