AI Agent Skills

MARS ships with three Agent Skills that teach AI coding agents (Claude Code, Cowork, and any tool supporting the open SKILL.md format) how to work with MARS-Curiosity. They live in the Skills/ folder of the repository:

Skill	What it does
mars-new-project	Scaffolds a new MARS server project — from the official MARSTemplate project group or from minimal bundled templates — and covers deployment (Windows service, ISAPI on IIS, Apache module, FastCGI, Linux daemon, HTTPS/SSL).
mars-development	Day-to-day MARS development: resources and REST attributes, parameter binding, JWT authentication and roles, FireDAC dataset publishing, Server-Sent Events, WebStencils templating, client components, configuration and serialization.
mars-mcp-server	Building MCP servers with MARS: exposing Delphi methods as tools for AI agents with [MCPTool], database tools via FireDAC, per-tool role authorization, Bearer and OAuth 2.1 authentication.

A skill is a folder with a SKILL.md file (instructions plus a description that tells the agent when to activate it) and optional references/ and assets/ files loaded on demand. Once installed, the agent picks the right skill automatically based on what you ask — you don't need to mention MARS explicitly if your code clearly uses it.

Installation in Claude Code (recommended: plugin)

The MARS repository is also a Claude plugin marketplace. From Claude Code:

/plugin marketplace add andrea-magni/MARS
/plugin install mars-curiosity@mars

All skills become available immediately and are updated whenever a new version is published. They can also be invoked explicitly as slash commands:

/mars-curiosity:mars-new-project
/mars-curiosity:mars-development
/mars-curiosity:mars-mcp-server

Installation with npx (any AI tool)

If you use other agentic tools besides Claude Code (Cursor, Codex, ...) or prefer not to use plugins, the skills CLI installs the same skills into any compatible tool with one command (requires Node.js, no permanent install):

npx skills add andrea-magni/MARS

Pick the skills and target tools when prompted (--all installs everything for every detected tool), and update later with npx skills update.

Manual installation (Claude Code / Cowork)

Alternatively, copy the skill folders into one of the standard skill locations:

• .claude/skills/ inside your project — project-level, shared with your team via git;
• ~/.claude/skills/ — personal, available in every project.

Example, from your project root (Windows):

git clone https://github.com/andrea-magni/MARS
xcopy /E /I MARS\Skills\mars-development .claude\skills\mars-development
xcopy /E /I MARS\Skills\mars-new-project .claude\skills\mars-new-project
xcopy /E /I MARS\Skills\mars-mcp-server .claude\skills\mars-mcp-server

Verify with /skills, or simply ask the agent to create a MARS server.

Installation in other AI agents

The skills are plain Markdown following the open Agent Skills format, so they are not tied to Claude:

• Agents with native skill support — install with npx skills add andrea-magni/MARS (see above) or place the skill folders in the agent's skills directory (the same SKILL.md layout is being adopted by a growing number of tools).

• Any other agent (Copilot, ...) — reference the skill files from your agent instructions file (AGENTS.md, .cursorrules, custom instructions, ...), for example:

  When working on MARS-Curiosity (Delphi REST) code, read and follow:
  - Skills/mars-development/SKILL.md (and its references/ folder)
  - Skills/mars-new-project/SKILL.md (when creating a new server)
  - Skills/mars-mcp-server/SKILL.md (when exposing tools to AI agents via MCP)

Since the instructions are self-contained Markdown, even a plain copy-paste into a system prompt works.

Usage examples

With the skills installed, prompts like these produce idiomatic MARS code without further guidance:

Scaffold a new server (activates mars-new-project):

Create a new MARS REST server called OrdersAPI, console host, listening on port 8090.

The agent generates the .dpr, Server.Ignition.pas, resource units and ini configuration from the bundled templates, with correct library paths and JWT backend selection.

Add an endpoint (activates mars-development):

Add a GET /customers/{id} endpoint returning a customer record as JSON.

[Path('customers'), Produces(TMediaType.APPLICATION_JSON)]
TCustomersResource = class
public
  [GET, Path('{id}')]
  function GetById([PathParam] id: Integer): TCustomer;
end;

Secure endpoints with JWT:

Protect the customers resource so only authenticated users with the 'admin' role can POST.

The agent knows about Token resources, [Context] Token: TMARSToken injection and the [RolesAllowed('admin')] attribute.

Publish a dataset:

Expose the EMPLOYEE table from my FireDAC connection as a REST endpoint.

The agent applies the FireDAC integration patterns from Skills/mars-development/references/firedac.md.

Expose tools to AI agents (activates mars-mcp-server):

Make my orders search callable from Claude via MCP, and restrict order cancellation to the 'manager' role.

The agent derives a TMCPResource (or TMCPDataResource for FireDAC-backed tools), annotates methods with [MCPTool]/[MCPParam] and applies per-tool [RolesAllowed] — see MCP Servers.

Deploy:

Turn this console server into a Windows service and prepare an ISAPI dll for IIS.

Deployment recipes come from Skills/mars-new-project/references/deployment.md.

Contributions

The skills are versioned with the library — improvements and corrections are welcome via Pull Request, like any other part of MARS.