Interview Prep AI — MCP Agent

An end-to-end interview coach that uses a real Model Context Protocol (MCP) server to ground an OpenAI agent in a candidate's resume and a target job description.

The project demonstrates MCP tool discovery, OpenAI Responses API tool calling, multi-step agent orchestration, document ingestion, deterministic evaluation logic, and a polished Streamlit UI.

What it does

• Upload a resume (PDF, DOCX, TXT, or MD)

• Upload or paste a job description

• Analyze matched skills and priority gaps

• Generate role-specific technical, behavioral, and system-design questions

• Evaluate answers with a STAR, relevance, specificity, and quantified-impact rubric

• Inspect every MCP call in an in-app activity trace

Related MCP server: shivonai-mcp

Architecture

flowchart LR
    U[Candidate] --> UI[Streamlit UI]
    UI --> A[OpenAI Responses agent]
    A <-->|tool schemas and calls| C[MCP client]
    C <-->|stdio| S[FastMCP server]
    S --> R[(Resume)]
    S --> J[(Job description)]
    S --> G[Gap analysis]
    S --> Q[Question generator]
    S --> E[Answer evaluator]

This is intentionally a genuine client/server MCP design. The LLM discovers JSON schemas from the MCP server, decides which tools to call, and receives each result through the Responses API function-calling loop.

MCP tools

Run locally

Prerequisites: Python 3.11+ and an OpenAI API key.

git clone https://github.com/YOUR_USERNAME/interview-prep-mcp.git
cd interview-prep-mcp
python -m venv .venv
source .venv/bin/activate       # Windows: .venv\Scripts\activate
pip install -e ".[dev]"
cp .env.example .env
# Add your OPENAI_API_KEY to .env
streamlit run app.py

The default is gpt-5.4-mini, chosen as a cost-conscious tool-calling model. Set OPENAI_MODEL in .env to use a different compatible model.

Test the MCP server

Run the unit tests without an API key:

pytest
ruff check .

Start the MCP server directly over stdio:

interview-prep-server

Privacy

Uploaded resume and JD text are stored only in .interview_prep/ on the local machine. That directory and .env are Git-ignored. Document text is sent to OpenAI only when the agent calls the relevant MCP tool during a workflow.

Design choices

• Auditable orchestration: the UI displays the exact MCP tools used for each answer.

• Testable core: skill comparison and answer scoring are deterministic; the LLM interprets and coaches rather than hiding all logic in a prompt.

• No invented experience: the system prompt requires resume evidence before claims about the candidate.

• Bounded agent loop: tool execution stops after eight rounds to prevent runaway calls.

Resume bullets

• Built an MCP-based Interview Preparation Agent using Python and OpenAI's Responses API, enabling an LLM to dynamically access resumes, job descriptions, and evaluation tools for personalized interview workflows.

• Implemented agentic tool-calling workflows for skill-gap analysis, targeted question generation, and rubric-based answer evaluation, with an auditable MCP activity trace.

Roadmap

• Persist separate interview sessions in SQLite

• Add voice answers and transcription

• Export a preparation report as PDF

• Add eval datasets for question quality and scoring consistency

• Deploy the MCP server with authenticated Streamable HTTP transport

References

• OpenAI function calling

• OpenAI model guide

• Build an MCP server

License

MIT