MCP Tools

Workflows support the Model Context Protocol (MCP) through the MCP Tools Provider operator node. This node connects deployed MCP servers to AI agents, giving agents autonomous access to external tools and integrations during execution.

How MCP Tools Work

MCP tools are not individual workflow nodes. Instead, they are accessed through a single operator node (mcp-tools-provider) that connects to an AI agent node. The agent autonomously decides which tools to call based on its instructions and the conversation context.

Architecture

[Trigger] → [AI Agent (react-agent)] → [Destination]
                    ↑
            [MCP Tools Provider]
            (connected via "tools" connector)
                    ↑
            [Deployed MCP Server(s)]

Key concepts:

• MCP Tools Provider: An operator node (type: mcp-tools-provider) that queries deployed MCP servers and provides their tool schemas to an agent
• Agent autonomy: The connected agent (e.g., react-agent, supervisor-agent) decides which tools to invoke based on its reasoning loop
• Tool connector: The MCP Tools Provider output connects to the agent's tools bottom connector, not the regular data flow

When to Use MCP Tools vs Native Nodes

Native workflow nodes are preferred over MCP equivalents when a direct integration exists. Use MCP tools when:

• No native workflow node exists for the service you need
• You need an AI agent to dynamically choose between multiple tools
• The task requires flexible, context-dependent tool selection

For example, if you need to write to PostgreSQL, use the native postgresql-dest node. If you need an agent to autonomously search the web, create GitHub issues, and send Slack messages based on context, use MCP tools connected to a react-agent.

MCP Tools Provider Node

Configuration

The mcp-tools-provider node has the following configuration fields:

Field	Type	Description
MCP Server	Select	Primary MCP server to provide tools from
Additional MCP Servers	Multi-select	Optional additional servers to combine tools from
Filter Tools	Tags	Only include tools matching these names (leave empty for all)
Cache Timeout	Number	How long to cache tool schemas in seconds (default: 300)

Inputs and Outputs

Outputs:

• tools - Array of tool definitions with name, description, parameters, and endpoint
• serverCount - Number of MCP servers queried
• toolCount - Total number of tools available

Connector

The output connector type is tools, which connects to an agent's tools input. It does not connect to regular data flow connectors.

Setting Up MCP Tools in a Workflow

Step 1: Add an Agent Node

Add a react-agent or supervisor-agent node to your workflow canvas.

Step 2: Add MCP Tools Provider

Add an mcp-tools-provider node to the canvas.

Step 3: Connect Tools to Agent

Connect the MCP Tools Provider's tools output to the agent's tools connector (the bottom connector on the agent node).

Step 4: Configure MCP Server

Select which deployed MCP server(s) the provider should query:

1. Click the MCP Tools Provider node
2. Select an MCP server from the dropdown
3. Optionally add additional servers
4. Optionally filter to specific tool names

Step 5: Configure the Agent

Configure the agent with a system prompt that describes when and how to use the available tools. The agent will see all tool schemas and autonomously decide which to call.

Available MCP Servers

The platform includes a library of containerized MCP servers that can be deployed. Each server exposes one or more tools following the MCP protocol standard.

Example server categories:

Category	Examples
Search	Brave Search, DuckDuckGo, Exa, arXiv
Development	GitHub, GitLab, DockerHub, CircleCI
Communication	Slack (via slack-dest node preferred), Microsoft Teams
Cloud	AWS Core, AWS Terraform, Azure AKS
Databases	Elasticsearch, Neo4j, CockroachDB
AI/ML	ElevenLabs, Whisper, Sequential Thinking
CRM & Productivity	Atlassian, Notion, HubSpot, Salesforce
Content	Contentful, Ghost, WordPress, Webflow
Analytics	Grafana, Metabase, PostHog
Security	SonarQube, Okta, Beagle Security

MCP Server Deployment

MCP servers must be deployed and running in the cluster before they can be selected in the MCP Tools Provider configuration. Each server runs as a containerized service with its own endpoint.

MCP Server Configuration

Each MCP server requires its own configuration, typically including credentials and endpoint details. MCP servers are configured separately from workflow nodes:

1. Server deployment: MCP servers are deployed as containerized services in the Kubernetes cluster
2. Credential management: Each server has its own API keys, tokens, or connection strings configured at deployment time
3. Tool discovery: When selected in an MCP Tools Provider, the provider queries the server's tool schemas at runtime

Example: Agent with MCP Tools

Web Research Agent

This workflow uses a react-agent with Brave Search MCP tools to research topics and return summaries.

Workflow structure:

webhook → react-agent → webhook-response
              ↑
      mcp-tools-provider
      (configured with brave-search MCP server)

Agent configuration:

• System prompt instructs the agent to search the web for the given query
• The agent autonomously calls the brave_search tool provided by the MCP server
• Results are synthesized by the agent's LLM and returned via webhook-response

Multi-Tool Agent

This workflow connects multiple MCP servers to a single agent:

webhook → react-agent → webhook-response
              ↑
      mcp-tools-provider
      (configured with github + brave-search + slack MCP servers)

The agent can autonomously:

1. Search the web for information (Brave Search)
2. Create GitHub issues or PRs (GitHub)
3. Send notifications (Slack)

The agent's system prompt determines when each tool is appropriate to call.

Native Nodes vs MCP Tools

Scenario	Recommended Approach
Write to PostgreSQL	Native postgresql-dest node
Read from MySQL	Native mysql-source node
Send email via SMTP	Native smtp node
Send Slack message	Native slack-dest node
Agent needs web search	MCP Tools Provider + Brave Search
Agent needs GitHub access	MCP Tools Provider + GitHub server
Agent decides between 5+ tools	MCP Tools Provider + multiple servers
Fixed data pipeline	Native nodes only (no MCP needed)

Best Practices

Tool Selection

1. Prefer native nodes for deterministic, fixed-step workflows
2. Use MCP tools when an agent needs to dynamically select tools
3. Filter tools in the provider to limit which tools the agent sees, reducing confusion
4. Combine servers when the agent needs cross-service capabilities

Agent Prompting

1. Be specific about when each tool should be used
2. Provide examples in the system prompt for complex tool interactions
3. Set boundaries on what the agent should and should not do with the tools

Performance

1. Cache tool schemas using the cache timeout setting to reduce MCP server queries
2. Limit tool count by filtering to relevant tools only
3. Use native nodes for high-throughput operations where agent reasoning overhead is unnecessary

Troubleshooting

MCP Server Not Available

• Verify the MCP server is deployed and running in the cluster
• Check that the server's health endpoint is accessible
• Review server logs for startup errors

Agent Not Calling Tools

• Verify the MCP Tools Provider is connected to the agent's tools connector (not the regular input)
• Check the agent's system prompt includes instructions for tool usage
• Ensure the MCP server is returning valid tool schemas

Tool Call Failures

• Check MCP server credentials are correctly configured
• Review the agent's execution spans for tool call error details
• Verify the external service (e.g., GitHub API, Brave Search) is accessible

Next Steps

• Learn about AI agents
• Test workflows
• Deploy workflows to production
• See workflow examples