What is the best way to use agents to build apps? Mother knows best.

November 4, 2026

First decide what the agent is for. An agent can write your app, or it can run inside your app, and each job has its own rules. Either way, check what already exists before you build.

A thin figure walks along a brick wall toward a glowing light bulb that stands on two large gears. — Walk toward the ideas...do it with machines. Do it with mothermachine.

Which job is your agent doing?

It writes your app. A coding agent such as Claude Code works in your repo, and you review what it changes.
It runs inside your app. Your users ask for something, and the agent does it for them.

A project can do both, so you may need both lists below.

What should you check first?

Look for software that already does the job. Your agent then changes settings instead of writing code, and that costs fewer tokens.

/mothermachine <what you need>

To start from working code. 117 of our 129 listings are adaptable. Each one names what to change, and 100 of them are Config only.
To use software as it is. 121 listings are operable, so an agent can drive them through an interface. Each one lists the jobs your agent can reach and the jobs it can't.

How do you get good work from a coding agent?

Keep a short instructions file. Put an AGENTS.md file at the root of your repo. List the commands that build and test the project, plus the rules the agent can't learn from the code. Leave out the file tree. The agent can read that for itself, and every extra line costs attention on every task. Claude Code reads CLAUDE.md, which can import AGENTS.md.
Give it one task at a time. Ask for one route, one component, or one integration. When something breaks, you know which change did it.
Give it a way to check its work. Without a test, a typecheck, or a mockup to compare against, the agent stops when the work looks done. Ask to see the command it ran and the output.
Put hard limits in tooling. A "do not touch" line in a prompt is a request, and the agent decides whether it applies. If a rule can't have exceptions, use a hook, a deny rule, or a sandbox.
Show it what you mean. Paste a screenshot or a mockup with the task. It saves you a paragraph about the layout.

We build this site the same way, with an AGENTS.md file, one branch per task, and checks that fail when a rule breaks.

How do you build an app an agent can run?

Let the model ask and your code act. The model returns a structured tool call. Your app checks it, adds the credentials, and runs it. The model never needs to see a key.
Give agents a standard way in. With the Model Context Protocol (MCP), a client asks your server which tools it has, then calls them. Any MCP client can use the same server. An API or a CLI works too. Across our 121 operable listings, the interface we record is MCP for 53, an API for 51, and a CLI for 13.
Handle sign-in in one place. Each user's keys and tokens need storing and refreshing, and that work repeats for every integration. MCP defines optional OAuth-based authorization for HTTP servers. Arcade is one service that runs the OAuth flows and stores and refreshes the tokens.
Let people say no. The MCP specification says a person should always be able to deny a tool call. Ask for confirmation before the agent spends money, sends a message, or deletes data.
Keep the agent and the screen in step. This matters when a person and the agent work on the same screen. A click updates the agent's context, and an agent action updates the page. AG-UI is an open protocol for this shared state. Its docs advise keeping sensitive data out of it.

What about low-code tools?

Relevance AI and Gumloop let you build agents without code, which is a quick way to test an idea. Both are hosted products, so check what you can take with you before you put real work on one.

Check a tool's status too. The GitHub repository for Flowise, an open-source visual builder, is archived.

What we do not cover yet

No single standard. Nobody sets a standard workflow for building with agents. This advice comes from vendor docs, open specifications, and our own work, and we have not measured it.
The tools named here. Claude Code, Arcade, Relevance AI, Gumloop, and AG-UI are not listings in the directory.

Frequently asked questions

What is the best way to use agents to build apps?

Decide whether the agent writes your app or runs inside it, because each job has its own rules. Then check what already exists, so your agent can configure software instead of writing it.

Is there one standard workflow for building with agents?

No. Nobody sets one. The advice here comes from vendor docs, open specifications, and our own work.

What goes in an AGENTS.md file?

The commands that build and test your project, plus the rules an agent can't learn from the code. Keep it short, and leave out anything the agent can find for itself.

Can a prompt stop an agent from changing a file?

Not every time. The agent decides whether a line in a prompt applies. Use a hook, a deny rule, or a sandbox for rules that can't have exceptions.

Should a model call third-party APIs directly?

No. Have the model return a tool call. Your app checks it, adds the credentials, and runs it.

Does my app need an MCP server?

No. An API or a CLI also lets an agent drive your software. MCP gives any MCP client the same way to list and call your tools.

How does my agent find software that already exists?

Install the skill, register the MCP server, or read the JSON files directly. For agents has the commands.

Sources

Best practices for Claude Code. Anthropic, read November 4, 2026.
Tool use with Claude. Anthropic, read November 4, 2026.
AGENTS.md. Agentic AI Foundation, read November 4, 2026.
Tools and Authorization. MCP specification, version 2026-07-28.