Laviya Agent Skills

Laviya Agent Skills is a repository of reusable orchestration assets for IDE and agent integrations. It provides a production-oriented MCP runtime, shared orchestration prompt assets, and client-specific artifacts for tools such as Cursor and Claude.

Packages

• laviya-agent-skills: reusable prompt, rule, and skill asset bundle for Node-based tooling
• laviya-mcp-server: stdio MCP runtime published from mcp/

What Is Included

• core/: shared orchestration prompt assets
• mcp/: TypeScript MCP runtime scaffold (laviya-mcp-server)
• cursor/: Cursor rule artifacts
• claude/: Claude skill artifacts (SKILL.md for orchestration, laviya-delivery-cycle/SKILL.md for the user-driven delivery cycle)
• docs/: setup, npm publishing, and end-user onboarding guides

Repository Structure

laviya-agent-skills/
  core/
  mcp/
    src/
    examples/
    README.md
  cursor/
  claude/
  docs/
    Setup.md
    MCPServerPublish.md
    InstallationAndUsage.md

Quick Start

cd mcp
npm install
npm run typecheck
npm run build
npm run dev

Required environment variable:

• LAVIYA_API_KEY

Install the Claude Code skills

Claude Code loads a skill only from a skills folder, one folder per skill named after the skill, with the file named SKILL.md:

The paths in the second column are where the files live in this repository, not where Claude Code looks for them. A file in ~/.claude/laviya-delivery-cycle/ (without skills) or directly in ~/.claude/ is ignored.

Windows (PowerShell, from the repository root):

New-Item -ItemType Directory -Force "$HOME\.claude\skills\laviya-orchestrator", "$HOME\.claude\skills\laviya-delivery-cycle" | Out-Null
Copy-Item mcp\examples\claude\SKILL.md "$HOME\.claude\skills\laviya-orchestrator\SKILL.md"
Copy-Item claude\laviya-delivery-cycle\SKILL.md "$HOME\.claude\skills\laviya-delivery-cycle\SKILL.md"

macOS / Linux (from the repository root):

mkdir -p ~/.claude/skills/laviya-orchestrator ~/.claude/skills/laviya-delivery-cycle
cp mcp/examples/claude/SKILL.md ~/.claude/skills/laviya-orchestrator/SKILL.md
cp claude/laviya-delivery-cycle/SKILL.md ~/.claude/skills/laviya-delivery-cycle/SKILL.md

Check: start a new Claude Code session and run /skills; laviya-orchestrator and laviya-delivery-cycle must be listed. Copy the files again after updating this repository. The Laviya MCP tools also work without the skills; the skills add the step-by-step rules.

Installable Asset Package

Install the reusable asset bundle:

npm install laviya-agent-skills

Example usage:

import { assets, resolveAssetPath } from "laviya-agent-skills";
import { readFileSync } from "node:fs";

const cursorRule = readFileSync(assets.cursorRule, "utf8");
const orchestratorPrompt = readFileSync(resolveAssetPath("orchestratorSystemPrompt"), "utf8");

Documentation Index

• Runtime architecture and implementation: mcp/README.md
• Developer setup guide: docs/Setup.md
• NPM publishing guide: docs/MCPServerPublish.md
• End-user installation and usage: docs/InstallationAndUsage.md
• Client-specific MCP setup (Codex, VS Code, Antigravity, Claude): docs/InstallationAndUsage.md

Runtime Overview

The MCP runtime follows a global-runtime plus project-local-config model:

• machine-level runtime installation
• repository-level project configuration (.laviya/project.json or .laviya.json)
• runtime-enforced lifecycle mechanics (validation, retries, idempotency, and structured logging)

Current MCP tool set includes both orchestration and local-direct task feed flows:

• laviya_help
• laviya_feed_task
• laviya_get_local_work_status
• laviya_cancel_local_work
• laviya_add_task_comment
• laviya_list_projects
• laviya_create_tasks
• laviya_start_delivery_run
• laviya_get_task
• laviya_upload_task_file
• laviya_update_task_description
• laviya_assign_task
• laviya_start_task_work
• laviya_complete_task_work
• laviya_get_my_work
• laviya_start_execution
• laviya_complete_execution
• laviya_report_token_usage
• laviya_diagnostics

All tools return API envelope JSON text in the shape: { HasFailed, Messages, Data }.

Work items may include a backend-defined ExecutionPolicy. Analysis and review policies are read-only: agents must not implement changes, and enforced policies require a truthful completion executionEvidence.

For full architecture details and file-level scaffolding, refer to mcp/README.md.