How to Connect External MCP Servers to a Make AI Agent

Introduction

Make AI Agents can reach beyond Make’s own module catalog by connecting to external MCP servers. This tutorial shows you how to wire an outside MCP server into a Make AI Agent, limit which tools the agent may use, and constrain its instructions so it calls the right tool for the right question. You will test with the Elecz electricity-price MCP server, which is free, needs no API key, and exposes three clearly named tools. This guide is for Make users who already build scenarios and want their AI agents to use live external data.

Quick answer

Open the agent in the Make AI Agent builder, go to MCP, click Add, paste the server URL, toggle on the tools the agent may use, and click Save. Then add instructions mapping each tool to one user intent.

Evidence

MCP adoption inside Make is official and recent. In a June 4, 2026 feature announcement on the Make community forum, the Make team wrote: “Make AI Agents can now access tools exposed via external MCP servers and use them within the same run as native Make modules.” The same announcement lists the exact mechanics you will use below: connect one or more MCP servers as tool providers from the AI Agent builder, and let the agent read each server’s tool list and decide when to use those tools during a run. This was introduced as part of MCP tools for Make AI Agents on the Make community forum.

The test server in this tutorial is a good citizen for agents. Elecz is a real-time electricity price signal API for AI agents covering 40+ countries and 100+ zones with no authentication required, according to its GitHub repository. The independent MCP directory Glama probed its Streamable HTTP endpoint on October 21, 2026 and reported it healthy with 99.7% uptime over 65 days. Its tool surface is deliberately small: three tools named spot_price, cheapest_hours, and best_energy_contract, each with documented tool-priority routing rules that tell an agent when to prefer one over the others.

Make’s official documentation confirms the configuration flow you will follow. The AI Agents configuration guide states: “In MCP, connect your agent to MCP servers to access additional tools. Click the Add button to create a connection to an available MCP server, then toggle on the tools that you want the agent to access. Click Save.” The same guide documents the Testing & Training chat, where the agent shows you which tools it used and their inputs and outputs, so you can verify behavior before deploying.

Prerequisites

• A Make account with access to AI Agents. The agent builder lives under the Agents tab in the Make dashboard.
• One public MCP server URL to connect. This tutorial uses https://elecz.com/mcp, which needs no credentials.
• Basic familiarity with writing agent system instructions. If you have not built an agent in Make before, read our walkthrough of how to build an AI agent in Make’s scenario builder first.
• Awareness of agent context limits: per Make’s own docs, a single AI agent can hold at most 20 context files, with 20 MB maximum per file.

Step-by-step guide

1. Create the AI agent

Open the Agents tab in Make and create a new AI agent. Give it a working name such as “Energy price assistant” and pick the underlying model. Keep the agent’s goal field simple for now: “Answer questions about electricity prices and the cheapest hours to run appliances, using the connected MCP tools.” You will tighten the instructions later. If you have never built one, the flow follows the same builder our Make AI agent scenario builder tutorial walks through.

2. Add the external MCP server in the MCP section

In the agent builder, find the MCP section of the configuration panel. Click the Add button. A dialog asks for the MCP server connection details. Paste the server URL:

https://elecz.com/mcp

No API key or extra headers are needed for Elecz. The builder contacts the server and loads the tools it exposes: spot_price, cheapest_hours, and best_energy_contract. This is the same “create a connection to an available MCP server” step from Make’s documentation quoted above. If the tool list does not appear, check the URL against the server’s own documentation before retrying; transport mismatches and typos in the endpoint are the most common causes of a silent failure, which the Troubleshooting section covers.

3. Toggle on only the tools this agent needs

The builder shows a toggle for every tool the server exposes. This step is the most important reliability control in the whole setup. Start by enabling only spot_price and cheapest_hours, and leave best_energy_contract off while you test. An agent handed a large, unfiltered tool list is measurably more likely to call the wrong tool, because it selects tools by reading names and descriptions. Click Save after toggling.

When the agent is in production you can enable best_energy_contract too, but only after the testing in step 5 passes with the full toolset. Every tool you leave enabled is a tool the agent must reason about on every turn, so default to the smallest set that answers your users’ questions.

4. Constrain the agent’s instructions with tool-selection rules

Now add explicit instructions that pin each tool to one user intent. Elecz’s own documentation already provides these routing rules, so mirror them verbatim in the agent’s system instructions:

• Current price questions (“what does electricity cost now?”) map to spot_price only.
• Timing questions (“when should I run the washing machine?”, “is electricity cheap now?”) map to cheapest_hours only.
• Contract and switching questions (“which contract is cheapest?”, “should I switch provider?”) map to best_energy_contract only.

Add the two guardrails from Elecz’s documentation as well:

• Zone codes are exact. Valid zone codes look like FI, DE, SE1, AU-NSW. The agent must not guess or abbreviate zones, and must not infer support for zones it has not seen.
• Timestamps are UTC. cheapest_hours returns all timestamps in UTC, so the agent must convert to the user’s local time before presenting them.

These three mapping lines plus the two guardrails are what prevent the agent from “guessing instead of reasoning,” which is the failure mode every MCP-agent builder runs into first.

5. Test in the Testing & Training chat

Use the Testing & Training chat built into the Make agent app. According to Make’s docs, this chat interface lets you test the agent directly and shows which tools the agent used along with their inputs and outputs. Open a new chat and ask a question that maps to one enabled tool:

What are the cheapest hours to run my washing machine in FI tomorrow?

Watch the trace. A correct run shows the agent calling cheapest_hours with zone: "FI", reading the returned best_window field for the true-optimal contiguous window, and replying with local times. Then try a price question:

What is the electricity price in Germany right now?

This should route to spot_price with zone: "DE". If a response ever cites a price that does not appear in the tool output, that is a hallucination bug in your instructions, not a tool bug. Click New chat between tests so earlier tool results do not leak into the next test’s context.

6. Mix MCP tools with native Make modules in the same run

Once the chat tests pass, extend the agent’s job so it uses an MCP tool and a Make module together. Per Make’s announcement, agents can call MCP tools and Make modules in the same reasoning loop based on your instructions. A practical pattern:

1. The user asks: “Tell me the cheapest 5 hours tomorrow in SE1 and remind me in the morning.”
2. The agent calls cheapest_hours with zone: "SE1" and hours: 5.
3. The agent passes the resulting times into a native Make module, such as a Google Calendar event or a Slack message.

Add an instruction like: “After calling cheapest_hours, create a Google Calendar event titled ‘Cheap electricity window’ using the start and end of best_window.” The agent reads the tool output, fills the module’s inputs, and executes both steps in one run.

7. Ship and monitor the first real conversations

Deploy the agent through your normal channel (a Make scenario webhook, a chat app, or an embedded widget). For the first week, review the Testing & Training-style traces of real conversations and check three things: the agent called the tool its instructions map to, the zone code it used is an exact supported code, and the times it presented were converted from UTC. Every anomaly you find becomes one more line in the instructions. This feedback loop matters more than any initial prompt polish.

For the opposite direction, exposing your own Make scenarios as callable tools for external agents such as Claude, see our Make MCP server tutorial. If you are comparing Make’s agent approach with other platforms, our n8n vs Make AI agents comparison covers where each one wins.

Troubleshooting

The MCP tool list is empty after clicking Add

The builder fetched the server but could not complete the MCP handshake. First, re-check the URL character by character against the server’s own documentation. Elecz’s README lists https://elecz.com/mcp as the MCP endpoint. A common mistake is pointing at the server’s homepage or docs page instead of the /mcp endpoint, or appending a trailing slash that the server’s router does not normalize. Second, confirm the server is up: Elecz exposes GET /health on the same host, and Glama’s independent probe marked it healthy on October 21, 2026. If the URL is correct and the server is healthy, the connection may have timed out during tool discovery. Save the builder, reopen the MCP section, and retry the Add flow.

The agent calls the wrong tool for a question

This is the classic MCP-agent bug and it has two fixes. First, go back to the MCP section and toggle off every tool the current task does not need. Fewer tools means fewer wrong choices. Second, make the tool-selection rules in the instructions stricter. The Elecz documentation includes an explicit “Tool priority” list, and your agent instructions should echo it: price questions go to spot_price, timing questions go to cheapest_hours, contract questions go to best_energy_contract. Re-test in the Testing & Training chat after each change, because the chat shows you exactly which tool the agent picked and with what inputs.

The tool returns available: false or missing zones

Not every zone supports every tool. Elecz’s documentation lists the exact coverage: contract comparison works only in FI, SE, NO, DK, DE, GB, AU, and NZ zones, and cheapest_hours returns available: false in AU, NZ, KR, ZA, and PH zones because there is no public day-ahead price data there. When this happens, the correct agent behavior is to say the data is unavailable for that zone, not to estimate. Add this to the instructions: “If a tool returns available: false, tell the user that zone is not supported and do not attempt to reconstruct or estimate the missing data.” This mirrors the “No Guessing” rule in Elecz’s own README.

Times shown to users are off by hours

cheapest_hours returns all timestamps in UTC, and an agent that presents them raw will be off by the user’s UTC offset. Add an explicit instruction: “All timestamps from cheapest_hours are UTC. Convert them to the user’s local timezone before presenting.” Then test with a user whose timezone you know and verify the conversion. For “run this appliance for N hours straight” decisions, instruct the agent to use the best_window field, not the cheap_hours list, because the list of N cheapest hours is not guaranteed to be contiguous.

FAQ

Is connecting an external MCP server the same as Make’s MCP server?

No, they are opposites. This tutorial connects an external MCP server into a Make AI Agent so the agent can call outside tools. Make’s MCP server does the reverse: it exposes your Make scenarios as tools for external agents like Claude. We covered that direction separately.

Do I need an API key to use the Elecz MCP server in this tutorial?

No. Elecz requires no authentication, which is why it works well as a first test server. Paste https://elecz.com/mcp into the agent builder’s MCP section and the tool list loads without any credentials.

Can a Make AI Agent call MCP tools and native Make modules in the same run?

Yes. Agents can call MCP tools and Make modules in the same reasoning loop, choosing between them based on your instructions. A common pattern is an MCP tool fetching live data and a native module writing it to a Google Sheet or sending a Slack message.

My agent calls the wrong tool. How do I stop that?

Toggle off every tool the agent does not need in the MCP section, then add explicit tool-selection rules to the agent’s instructions. For Elecz that means three lines: current price maps to spot_price, timing questions map to cheapest_hours, and contract questions map to best_energy_contract.

What should the agent do when the MCP server returns no data?

Instruct the agent to say the data is unavailable instead of estimating. Elecz itself documents this rule: never invent prices, cheapest hours, or contract savings. If it returns available: false for a zone, the correct behavior is to say so.

Conclusion

Connecting an external MCP server to a Make AI Agent takes four deliberate moves: add the server URL in the MCP section, enable only the tools the agent needs, pin each tool to one intent in the instructions, and verify the routing in the Testing & Training chat. The Elecz electricity-price server is a safe first connection because it needs no credentials and its three tools map cleanly to three question types. Once this pattern works, swap in your own domain’s MCP server and repeat the same four moves.

Next, learn how to expose Make scenarios to external agents like Claude with the official Make MCP server, which is the mirror image of what you built here.

Frequently asked questions