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/

• Remote MCP endpoint https://api.laviya.app/api/mcp (OAuth), registered from mcp-remote/server.json

Related MCP server: production-grade-mcp-agentic-system

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, laviya-prd-intake/SKILL.md for turning a PRD into Laviya work)

• docs/: setup, npm publishing, and end-user onboarding guides (docs/examples/: an example PRD)

Repository Structure

laviya-agent-skills/
  core/
  mcp/
    src/
    examples/
    README.md
  mcp-remote/
    server.json
  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", "$HOME\.claude\skills\laviya-prd-intake" | Out-Null
Copy-Item mcp\examples\claude\SKILL.md "$HOME\.claude\skills\laviya-orchestrator\SKILL.md"
Copy-Item claude\laviya-delivery-cycle\SKILL.md "$HOME\.claude\skills\laviya-delivery-cycle\SKILL.md"
Copy-Item claude\laviya-prd-intake\SKILL.md "$HOME\.claude\skills\laviya-prd-intake\SKILL.md"

macOS / Linux (from the repository root):

mkdir -p ~/.claude/skills/laviya-orchestrator ~/.claude/skills/laviya-delivery-cycle ~/.claude/skills/laviya-prd-intake
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
cp claude/laviya-prd-intake/SKILL.md ~/.claude/skills/laviya-prd-intake/SKILL.md

Check: start a new Claude Code session and run /skills; laviya-orchestrator, laviya-delivery-cycle and laviya-prd-intake 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 + MCP Registry publishing guide (stdio and remote entries, GitHub Copilot listing): 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_list_tasks

• laviya_create_tasks

• laviya_start_delivery_run

• laviya_get_task

• laviya_upload_task_file

• laviya_update_task_description

• laviya_assign_task

• laviya_list_sprints

• laviya_create_sprint

• laviya_assign_tasks_to_sprint

• laviya_list_task_links

• laviya_add_task_link

• laviya_remove_task_link

• laviya_list_test_cases

• laviya_create_test_cases

• laviya_report_test_result

• laviya_list_wikis

• laviya_get_wiki

• laviya_create_wiki

• laviya_update_wiki

• 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.