MCP 服务器

interview-prep-mcp

An open-source Python MCP server for AI-assisted interview preparation.

README

interview-prep-mcp

An open-source Python MCP server for AI-assisted interview preparation. Runs in any MCP-compatible host — Claude Desktop, Cursor, Google Antigravity, Windsurf — and defaults to Google Gemini with optional OpenRouter for cross-provider model selection.

What it does

interview-prep-mcp exposes 17 tools that cover the full interview-prep loop:

Parse a job posting (URL or pasted text) and a CV (PDF or markdown).
Persist company and interviewer research collected by the host's other MCPs (e.g. firecrawl, exa) — this server doesn't scrape itself.
Generate personalized interview questions using the active LLM.
Submit practice answers and get scored evaluations.
Analyze weak areas across your practice history.

It also exposes:

A interview-prep://interviews/{id} resource bundling the full prep state for one session.
Two MCP prompts: daily_prep and mock_interview.

Install

The fastest way — no global pip install:

uvx interview-prep-mcp

Claude Desktop (`claude_desktop_config.json`)

{
  "mcpServers": {
    "interview-prep": {
      "command": "uvx",
      "args": ["interview-prep-mcp"],
      "env": {
        "GEMINI_API_KEY": "AIza...",
        "OPENROUTER_API_KEY": "sk-or-..."
      }
    }
  }
}

OPENROUTER_API_KEY is optional — leave it out if you only want to use Gemini.

Google Antigravity

Antigravity uses the same mcpServers JSON in its MCP settings panel. Paste the same block as above. Antigravity ships with Gemini already configured at the host level, so you'll typically only need to add the MCP entry itself.

Cursor (`~/.cursor/mcp.json`)

Same mcpServers block as Claude Desktop.

Windsurf and other MCP-compatible hosts

Same mcpServers block; just paste it into whichever config path the host documents.

Required env

Variable	Required?	Purpose
`GEMINI_API_KEY`	Yes (default)	Direct Gemini API access.
`OPENROUTER_API_KEY`	No	Enables `set_model` to switch to any of 200+ OpenRouter models.

You can run with only one of them set. If neither is set, the server still starts — context tools (parsing, storage) work; smart tools will return a clear "no LLM provider configured" error when invoked.

Recommended companion MCPs

To research companies and interviewers, install these alongside:

firecrawl-mcp — scrape pages.
exa-mcp-server — web search.

The host orchestrates all three. interview-prep-mcp exposes save_research_note for the host to persist whatever those companion MCPs returned.

Quick start (5 minutes)

Set GEMINI_API_KEY and install the server in your host (see snippets above).
In a fresh chat: parse_job_posting("https://example.com/jobs/123").
parse_cv("/path/to/cv.pdf") — first CV becomes active automatically.
start_interview_prep("Acme", "Senior Backend Engineer", job_id=1, cv_id=1).
(Optional) Ask Claude to research the company via firecrawl/exa and call save_research_note(interview_id=1, kind="company", content="...").
generate_questions(interview_id=1, count=10, types=["technical","behavioral"]).
Answer questions; each answer goes to submit_practice_answer.
After several answers: analyze_weak_areas(interview_id=1).

Tool reference

Tool	LLM?	Description
`parse_job_posting(source_or_text, is_html?)`	no	URL or text → structured JobPosting.
`parse_cv(file_path)`	no	PDF/markdown → CVProfile. First CV is set active.
`list_cv_profiles()`	no	All CVs with their `is_active` flag.
`set_active_cv(cv_id)`	no	Switch which CV other tools default to.
`match_cv_to_job(job_id, cv_id?)`	yes	Score + strengths + gaps.
`start_interview_prep(company, role, job_id?, cv_id?)`	no	New prep session.
`list_interviews()`	no	Active prep sessions.
`save_research_note(interview_id, kind, content, subject?, source_url?)`	no	Persist markdown research. `kind ∈ {company, interviewer, other}`.
`get_research(interview_id)`	no	Grouped by kind.
`generate_questions(interview_id, count, types)`	yes	Personalized Qs using job + cv + research.
`list_questions(interview_id)`	no	Previously generated.
`submit_practice_answer(question_id, answer_text)`	yes	Returns evaluation.
`get_practice_history(interview_id)`	no	Past attempts.
`analyze_weak_areas(interview_id)`	yes	Recurring gaps + recommended topics.
`list_models(provider?, ...filters)`	no	Gemini (curated) + OpenRouter (24h cached).
`set_model(model_id, provider?)`	no	Persist active model.
`get_active_model()`	no	Current selection.

Architecture

MCP Host
├── firecrawl-mcp ──────┐
├── exa-mcp ────────────┤
└── interview-prep-mcp ─┘
        │
        ├── Tools layer
        ├── Service layer (parsers + storage + LLM router)
        ├── SQLite (~/.interview-prep/db.sqlite)
        ├── Config (~/.interview-prep/config.json)
        └── LLM router ──┬── Gemini (default)
                         └── OpenRouter (optional)

The MCP composes with companion MCPs — it does not scrape itself. The host's Claude calls firecrawl/exa, then feeds findings into save_research_note.

Contributing

git clone https://github.com/shenmali/Interview-MCP-First
cd interview-prep-mcp
python3.11 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pytest
ruff check .
mypy interview_prep_mcp

CI runs ruff + mypy + pytest on every PR.