Creating an agent leveraging documentation

In this example, we're going to crawl a documentation website, store the content in a vector database and leverage it to anwser questions about it.

Create the vector database

We're going to use the doc2vec project to crawl the MCP documentation website and store the content in a sqlite-vec database.

Clone the doc2vec repo:

git clone https://github.com/kagent-dev/doc2vec.git

Install the dependencies:

cd doc2vec
npm install

Update the configuration file to crawl the MCP documentation sqlite:

cat <<EOF > config.yaml
sources:
  - type: website
    product_name: 'mcp'
    version: 'latest'
    url: 'https://modelcontextprotocol.io/'
    max_size: 1048576
    database_config:
      type: 'sqlite'
      params:
        db_path: './mcp.db'
EOF

Launch the process:

npm start

It will take several minutes to complete.

Create an MCP server to use the MCP documentation database

Now that we have the database, we can create an MCP server to leverage it.

The goal is to deploy it on Kubernetes to allow our agent to use it.

To simplify the process, we're going to store the database in the Docker image.

Let's build and push the Docker image:

cd mcp
cp ../mcp.db .
docker build -t <your docker repository> .
docker push <your docker repository> .

Deploy the MCP server

Create a Secret for the OpenAI API Key which is going to be used to create the embeddings:

kubectl create secret generic mcp-secrets \
  --from-literal=OPENAI_API_KEY=<your_openai_api_key>

Create a ConfigMap for the Database Configuration:

kubectl create configmap mcp-config \
  --from-literal=SQLITE_DB_DIR=/data \
  --from-literal=PORT=3001

Create a file named deployment.yaml:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: mcp-sqlite-vec
  labels:
    app: mcp-sqlite-vec
spec:
  replicas: 1
  selector:
    matchLabels:
      app: mcp-sqlite-vec
  template:
    metadata:
      labels:
        app: mcp-sqlite-vec
    spec:
      containers:
      - name: mcp-sqlite-vec
        image: <your docker repository>
        imagePullPolicy: Always
        ports:
        - containerPort: 3001
        env:
        - name: OPENAI_API_KEY
          valueFrom:
            secretKeyRef:
              name: mcp-secrets
              key: OPENAI_API_KEY
        - name: SQLITE_DB_DIR
          valueFrom:
            configMapKeyRef:
              name: mcp-config
              key: SQLITE_DB_DIR
        - name: PORT
          valueFrom:
            configMapKeyRef:
              name: mcp-config
              key: PORT

Apply it:

kubectl apply -f deployment.yaml

Create a file named service.yaml:

apiVersion: v1
kind: Service
metadata:
  name: mcp-sqlite-vec
spec:
  selector:
    app: mcp-sqlite-vec
  ports:
  - port: 3001
    targetPort: 3001
  type: ClusterIP

Apply it:

kubectl apply -f service.yaml

Use the MCP server in kagent

In the kagent UI, click on Tools and then on Manage tool servers.

Click on Add Server.

Call it sqlite-vec.

Select URL and use http://mcp-sqlite-vec.default:3001/sse.

Click on Add Server.

After refreshing the page, you should see the query-documentation tool being discovered.

Create the MCP agent

Click on New Agent

Use the following information:

• Agent Name: mcp-agent
• Description: The MCP agent is answering questions about MCP, using the MCP documentation
• Agent Instructions: Use your tool to answer any question about the Model Context Protocol (MCP). Use mcp for the product and latest for the version.

Click on Add Tools.

Select the query-documentation tool.

Congratulations, the agent is ready !

Use the MCP agent

Click on the agent and ask a question.

For example, How can you build an MCP server using sse ?

As you can see, the agent has used the query-documentation tool to answer the question.