Coding agents

What a coding agent (Claude Code, Codex, Cursor, Gemini CLI, Elyra Workspace…) needs to work well on an Elyra app, the way Laravel Boost does it for Laravel: guidelines it reads up front, skills it loads when a task needs them, and an MCP server that answers questions about the app.

This is about agents writing the app. An agent using the app, calling its commands as tools, is the app as an MCP server.

In two lines

rata ai:install     # AGENTS.md block, skills, and `rata mcp dev` in .mcp.json
rata ai:update      # refresh them after upgrading rata

What ai:install writes

• Guidelines in AGENTS.md, between <elyra-framework-guidelines> and </elyra-framework-guidelines>: commands and abilities, middleware, the container, models and migrations, live queries, rata codegen after a type changes, which rata generator to use, and the checks. Only the block is rata's. What you write around it stays, and ai:update replaces only the block. A CLAUDE.md gets the same block, unless it imports AGENTS.md (@AGENTS.md).

• Skills in .agents/skills/ and .claude/skills/, in the Agent Skills format, loaded only when they apply:

  Skill	For
  elyra-commands	commands, Ctx, validation, abilities, middleware, events, TestApp
  elyra-models-and-migrations	#[derive(Model)], the query builder, relations, casts, factories, SQL and Rust migrations
  elyra-resources	rata make:resource, the registries, changing a generated resource
  elyra-live-queries	#[command(live)], live.*, depends_on / invalidate, testing
  elyra-laravel-backend	Backend, Auth, Laravel's answers, broadcasting, the outbox, notifications

  A skill file ends with rata's marker. One without it is yours: rata says so and leaves it, so delete it to get the framework's back.

• The elyra-framework server in .mcp.json: rata mcp dev, next to the servers already there (Laravel Boost's, say). Claude Code reads the file, and Elyra Workspace gives its servers to every agent once you allow it.

ai:update refreshes only what's there: a removed skill or server stays removed. Run it after upgrading rata, so the guidelines match the framework.

rata mcp dev

An MCP server over stdio for the agent working on the code. Every tool only reads: nothing is migrated, run or written, and a SQLite database is opened read-only.

• The code is read with syn, as make:resource reads a model, so app_info needs no build.
• The log is the one LogProvider writes (to_app_dir("My App") → <config dir>/my-app/app.log, or to_path(..)), found from the source, then from the bundle name and the crate. Rotated app.log.1 is read too. Give path when it's elsewhere.
• The database comes from [database].url or DATABASE_URL, as for rata migrate. A SQLite file that isn't there isn't created.
• Protocol: clients that open with initialize (2025-11-25, 2025-06-18, 2025-03-26), and 2026-07-28 requests carrying their version in _meta.

The Laravel half

When the app is the desktop half of a Laravel app (a Laravel backend), the Composer package elyracode/elyra-backend in this repository gives the Laravel app's agents the same knowledge from the other side. Laravel Boost picks up its guidelines and its elyra-desktop-backend skill: what the desktop expects of the API (the Sanctum token routes, 422s per field, the status codes, Idempotency-Key, If-Unmodified-Since, broadcasting, notifications), and that changing an API resource means changing the desktop's types and running rata codegen. The package also has the routes and the idempotency middleware.

A project with both halves then gives an agent both servers: Boost's laravel-boost for the API, and elyra-framework for the desktop.

Related

• AI agents (MCP) — the app's own commands as tools
• Ratatosk CLI · A Laravel backend