MCP and AI agents · Zeyn docs

Zeyn is built for agents on two sides: agents that write Zeyn apps, and agents that use a running app on behalf of a user.

Agents that write apps

The project an agent works in carries its own rules and tools:

AGENTS.md and CLAUDE.md list the commands, the project structure, the rules of the language and a framework index.
.mcp.json registers zeyn mcp, an MCP server on stdio with these tools:

Tool	Does
`zeyn_docs`	Framework reference by topic: `rules`, `language`, `protocol`, `architecture`
`zeyn_describe`	Features, screens, routes, access, APIs, problems and missing translations of the app
`zeyn_check`	Compiles and reports problems; `fix: true` first rewrites renamed APIs
`zeyn_test`	Runs `cargo test` and returns the end of the output
`zeyn_new_feature`	Creates a feature folder and registers it

The compiler is the main guard. Every rule it can enforce has a compile_fail test and a #[diagnostic::on_unimplemented] message written for an agent to act on, such as:

error[E0277]: `min()` and `max()` cannot be used on a `bool` column
  = note: `min()` and `max()` work on i32, i64, Money, Date, Time, DateTime, LocalDateTime and their Option forms

ZEYN_DESCRIBE=1 cargo run prints the same description zeyn_describe returns and exits.

Agents that use an app

App::new(t!("app.name")).agents(Agents::read_only())

With agents, the app's screens become tools:

one search tool for each grid,
one save tool for each form, with the form's fields and rules as its schema (only with Agents::read_write()).

Agents reach them in two ways:

In the browser through WebMCP (document.modelContext.registerTool), when the browser supports it. The tools of the current screen are registered as it opens and removed when it closes.
On the server at POST /_zeyn/mcp (JSON-RPC). It needs an Authorization header, so a browser session alone can never drive it.

The same rules as people

An agent gets exactly the permissions of the user it acts for. Tools for screens and actions the user may not open are not listed.
Calls run the guards, interceptors, validation and rate limits of a normal call.
Audit records mark calls made by agents, so a reviewer can tell them apart.
Agents::read_only() lets agents search but not save. Choose read_write() only for apps where an agent should change data.