Build MCP tools with AI: describe it, review it, publish it
Describe a tool in plain words and Ada writes, builds and tests it in Mesh Builder. You allow its hosts and keys, then publish. Tips for a description that works.
In this post
- Start a build in Mesh Builder
- What Ada does, step by step
- The one step only you can take
- How to write a good description
You don't need to write code to give an AI assistant a new tool. Say what the tool should do, link the API's docs, and Ada, the AI in Mesh Builder, writes it, builds it and tests it. You check it and publish it.
Start a build#
Open Mesh Builder in the console and choose New build. Fill in two boxes:
| Box | Example |
|---|---|
| What should it do? | Refund a Stripe charge, list refunds and look one up. Never refund more than the charge. |
| Link to the API's docs | docs.stripe.com/api/refunds (up to 5 links) |
A build usually takes a minute or two, and you watch it live. Each step is on the left, Ada's work is in the middle, and each tool appears on the right as it's written and tested. You can type to her while she works. She reads your message before her next step.
What Ada does#
| Step | What happens |
|---|---|
| 1. Read and plan | She reads your links and says which tools she will write |
| 2. Write | She writes each tool, preferring ones that need no code |
| 3. Build and test | The draft is built and its tests run |
| 4. Your turn | You allow the hosts it reaches and add the keys it needs |
| 5. Publish | The draft goes live after the same checks a push runs |
If a test fails, tell Ada what to fix, or choose Try a call and run the tool with your own inputs.
The step only you can take#
Ada can't give a tool network access, add keys or publish. Each draft lists the hosts it reaches and the keys it needs. You choose Allow for each host and Store key to paste a key's value into the project's vault.
Write a good description#
Ada works best with what you'd tell a colleague:
- What it's for, in one sentence. This becomes the tool's description, which the AI model reads when it decides whether to call it.
- The inputs, with an example value for each.
- What to return, field by field. Fewer fields cost fewer tokens on every call.
- The docs, as links. Ada uses only what the docs say and won't invent endpoints.
- What to avoid, such as "never delete" or "read only".
After it's live#
Keep chatting to Ada about the tool. Ask how to connect a client, what a tool does, or ask her to change it. A change comes back as a new draft for you to review.
Prefer your own agent? Connect Claude Code, Cursor or any MCP client to your project and ask it for the tool. It proposes a draft that you review in the same way. The Build with AI guide has the details, and Turn your API into MCP tools from GitHub covers building from your own code.
Explore MCP guides by task: creation, hosting, connections and security.
