MCP Resume Assistant (FastMCP + LangGraph + Streamlit)

This project is a production-style MCP demo for resume bullet rewriting. It combines:

• mcp-server: a FastMCP tool server (resume_bullet_tool)
• mcp-host: a FastAPI + LangGraph agent host that calls MCP tools and Gemini
• streamlit: a lightweight web UI that calls the host API

The repository includes Dockerfiles and Kubernetes manifests for deployment on GKE.

1) Features

• Rewrite weak resume bullets into stronger, impact-oriented wording
• Return one best variant plus two alternative variants
• Score bullet quality (0-10) with specific improvement tips
• Encourage metrics and measurable outcomes without fabricating numbers
• Support role-targeted phrasing (for example, Data Engineer or Software Engineer)
• Provide end-to-end deployability with Docker and Kubernetes

2) Project Structure

.
├── mcp-server/                 # FastMCP tool service (port 9000)
│   ├── mcp_tool_server.py
│   ├── requirements.txt
│   └── Dockerfile
├── mcp-host/                   # LangGraph + FastAPI host (port 8000)
│   ├── src/
│   │   ├── agent/graph.py
│   │   └── api/apiserver.py
│   ├── requirements.txt
│   └── Dockerfile
├── streamlit/                  # UI (port 8080 in container)
│   ├── app.py
│   ├── requirements.txt
│   └── Dockerfile
└── k8s/                        # Kubernetes manifests (namespace: extraquiz)

3) How It Works

Request flow:

1. User enters prompt in Streamlit
2. Streamlit sends POST /invoke to mcp-host
3. mcp-host LangGraph agent calls MCP tools from mcp-server
4. Agent returns rewritten bullet suggestions + score info
5. Streamlit displays JSON response

Internal service URLs in K8s:

• mcp-server: http://mcp-server:9000/mcp
• mcp-host: http://mcp-host:8000

4) Prerequisites

• Python 3.10+ (3.11 recommended for host/server)
• Docker
• Google Cloud SDK (gcloud)
• kubectl
• A GCP project with GKE enabled
• A valid GOOGLE_API_KEY (Google AI Studio / Gemini)

5) Local Run (Without Kubernetes)

4.1 Start MCP server

cd mcp-server
pip install -r requirements.txt
fastmcp run mcp_tool_server.py:mcp --transport http --host 0.0.0.0 --port 9000

4.2 Start MCP host

In a new terminal:

cd mcp-host
pip install -r requirements.txt
export GOOGLE_API_KEY="YOUR_KEY"
export MCP_TOOL_URL="http://localhost:9000/mcp"
export GEMINI_MODEL="gemini-3-flash-preview"
uvicorn api.apiserver:app --host 0.0.0.0 --port 8000 --app-dir src

Health check:

curl http://localhost:8000/ok

4.3 Start Streamlit UI

In a third terminal:

cd streamlit
pip install -r requirements.txt
export MCP_CLIENT_API_URL="http://localhost:8000"
streamlit run app.py --server.port 8080

Then open http://localhost:8080.

6) Deploy to GKE

5.1 Create cluster

gcloud container clusters create extraquiz \
  --machine-type e2-standard-4 \
  --num-nodes 2 \
  --zone us-central1-a \
  --cluster-version latest

5.2 Authenticate

gcloud auth login
gcloud auth configure-docker

5.3 Build and push images

Update image registry/project if needed, then run:

gcloud builds submit ./mcp-server --tag gcr.io/mcppp/mcp-server:1.0
gcloud builds submit ./mcp-host --tag gcr.io/mcppp/mcp-host:1.0
gcloud builds submit ./streamlit --tag gcr.io/mcppp/streamlit:1.0

5.4 Create namespace and secret

kubectl apply -f k8s/namespace.yaml

Before applying k8s/secrets-template.yaml, replace GOOGLE_API_KEY with your real key:

kubectl apply -f k8s/secrets-template.yaml

5.5 Deploy all workloads

kubectl apply -f k8s/mcp-server-deployment.yaml
kubectl apply -f k8s/mcp-server-service.yaml
kubectl apply -f k8s/mcp-host-deployment.yaml
kubectl apply -f k8s/mcp-host-service.yaml
kubectl apply -f k8s/streamlit-deployment.yaml
kubectl apply -f k8s/streamlit-service.yaml

5.6 Verify

kubectl get pods -n extraquiz
kubectl get svc -n extraquiz

When streamlit service gets an external IP (LoadBalancer), open it in browser.

7) API Reference

Host health endpoint

• GET /ok
• Response:

{"status":"ok"}

Host invoke endpoint

• POST /invoke
• Request body:

{"query":"Rewrite this resume bullet for a Data Engineer role: worked on ETL pipelines"}

• Response body:

{"output":"...agent final response..."}

8) Key Environment Variables

mcp-host

• GOOGLE_API_KEY (required)
• MCP_TOOL_URL (default: http://localhost:9000/mcp)
• GEMINI_MODEL (default: gemini-3-flash-preview)

streamlit

• MCP_CLIENT_API_URL (default: http://localhost:8000)

9) Notes

• The tool server uses a heuristic scoring approach for bullet quality.
• If your bullet has no metrics, the system will encourage placeholders instead of fabricating numbers.
• For production use, consider adding authentication, rate limiting, centralized logging, and CI/CD.