AI & MCP
What an AI coding agent is
An AI coding agent reads the repository, edits files and runs commands in a loop until the tests pass: it is not autocomplete. Here is what it needs.
Key points
- A coding agent runs a loop: it reads, edits, runs commands and corrects.
- The difference from autocomplete is access to the repository and the tools, not the model.
- Every diff is code entering the project: review stays with you.
An AI coding agent is a program that reads the repository, edits files and runs commands in a loop until a test or a request is satisfied. It is not autocomplete: it chooses the steps and calls tools.
The loop
It reads, proposes an edit, applies it, runs a command, reads the output and corrects. The loop ends when the command passes or when the agent asks for a decision. Quality depends on how much context it can read.
What separates it from autocomplete
Autocomplete suggests a line at the cursor. An agent opens files you never opened, finds a symbol across the repository and runs the test suite. The price is review: every diff is code entering your project.
Connecting a catalogue to the agent
claude mcp add --transport http eneui https://ui.eneawork.it/mcpThe command registers the EneUI MCP endpoint as an HTTP server. From then on the agent searches a component and reads its licence before proposing a file; the protocol is described in what MCP is. The full path for Claude Code, Cursor and VS Code is in MCP on Claude Code, Cursor and VS Code.
What it needs to avoid mistakes
A repository that builds, fast tests and a catalogue with explicit licences. With those three ingredients the agent checks its own assumptions; without them it produces plausible code nobody verified. A readable record is AI chat input.
The catalogue as context
EneUI exposes 111,985 components and 1,556 blocks with the upstream licence. The agent can cite the source instead of inventing an API; the downloaded file stays your code, as in the copy-paste or headless model.
The limitation
An agent runs commands on your disk. A destructive test, a migration or a badly formed rm reaches the filesystem without asking, unless the client enforces explicit approval. Read the diff and restrict the allowed commands.
Where to continue
For the client setup read MCP on Claude Code, Cursor and VS Code; for review rules on generated code see the accessibility checklist.
Frequently asked questions
Is a coding agent a chatbot?
No. A chatbot answers with text; an agent reads files, applies edits and runs commands in your repository.
Does it need a paid API?
Yes, almost always. The agent calls a hosted model and pays per token; free clients carry usage limits, not the absence of a model.
Can the agent delete files?
Yes, if it has access to the filesystem and to commands. Restrict the allowed commands and read the diff before you accept it.
What is MCP here?
It is the protocol an agent uses to call external tools, for example a search in a component catalogue. Without MCP the agent sees only the repository.
Do I own the generated code?
It depends on the licence of the source and on the model contract. Copying an MIT component yields MIT code; a model with its own clauses can add conditions.