Nadhebe
tutorials

Authoring Custom Agent Skills and Packaging MCP Bundles (.mcpb) for IDE Integration

A step-by-step tutorial on building custom SKILL.md playbooks and packaging zero-dependency Model Context Protocol Bundles (.mcpb) for AI IDEs.

Nadhebe Editorial Team Nadhebe Editorial Team
· · 2 min read
GPU Lab Verified
Retro-tech editorial graphic showing packaged archive bundles and SKILL.md manifests on a soft mint background
On this page

Authoring Custom Agent Skills and Packaging MCP Bundles (.mcpb) for IDE Integration

The standardization of Agent Skills alongside Model Context Protocol Bundles (.mcpb) simplifies how AI engineering capabilities are distributed across development environments.

Historically, installing local MCP tools required end users to manually provision Node.js or Python runtimes and edit hidden JSON configuration files. The .mcpb archive specification packages local stdio tools together with embedded execution runtimes, enabling single-click zero-dependency installation.


Agent Skills (SKILL.md) vs MCP Bundles (.mcpb)

 ┌────────────────────────────────────────────────────────────────────────┐
 │                      AGENT TOOLING DISTRIBUTION                        │
 └───────────────────────────────────┬────────────────────────────────────┘

           ┌─────────────────────────┴─────────────────────────┐
           ▼                                                   ▼
┌──────────────────────────────┐                    ┌──────────────────────────────┐
│     SKILL.MD PLAYBOOKS       │                    │      MCP BUNDLES (.MCPB)     │
├──────────────────────────────┤                    ├──────────────────────────────┤
│ • Instructions & Prompts     │                    │ • Binary Runtime Archive     │
│ • YAML Frontmatter Metadata  │                    │ • Manifest Definition        │
│ • No Runtime Binary Needed   │                    │ • Single-Click Installation  │
└──────────────────────────────┘                    └──────────────────────────────┘
LayerPrimary ResponsibilityFile ManifestTarget User Experience
Agent SkillDefines multi-step workflows & promptsskills/<name>/SKILL.mdAuto-discovered domain expertise
MCP BundleDistributes executable binary toolsbundle.json in .mcpbZero-dependency 1-click install

Step 1: Authoring a Custom SKILL.md Playbook

Create a skill directory at .claude/skills/db-migration/SKILL.md:

---
name: database-migration
description: Automated Prisma & PostgreSQL schema migration playbook with rollback checks
---

# Database Migration Skill Playbook

When the user asks to run or generate database migrations, follow these steps:

1. **Schema Check**: Inspect `prisma/schema.prisma` for uncommitted structural changes.
2. **Dry Run**: Execute `npx prisma migrate dev --create-only` to generate sql file without applying.
3. **Audit**: Inspect the generated SQL script for non-reversible drops (`DROP TABLE` / `DROP COLUMN`).
4. **Apply**: Execute `npx prisma migrate deploy` after user confirmation.

Step 2: Packaging an MCP Bundle (.mcpb)

Build an .mcpb package using build-mcpb:

# Install the MCP Bundle CLI Tool
$ npm install -g @modelcontextprotocol/build-mcpb

# Compile the local MCP server into a single portable bundle
$ build-mcpb --input ./my-mcp-server --output db-tools.mcpb

Once generated, double-clicking db-tools.mcpb opens your desktop AI client and installs the MCP tool bundle instantly with zero manual JSON editing.

Frequently asked questions

What problem do MCP Bundles (.mcpb) solve for developers?

.mcpb archives package local MCP tools directly with embedded execution runtimes, allowing developers to install tools without manually configuring Node.js or Python environments.

How do AI assistants discover custom agent skills?

Assistants inspect repository root folders, `.claude/skills/` directories, or installed plugin registries for valid `SKILL.md` manifests.

Are MCP Bundles compatible across different AI IDE platforms?

Yes, standard `.mcpb` archives conform to the open Model Context Protocol specification, running across supporting desktop environments.

Sources & references

  1. [1]Model Context Protocol Bundle Specification
Nadhebe Editorial Team

Nadhebe Editorial Team

The collective editorial desk, technical writers, and hardware validation engineers at Nadhebe. All publications undergo multi-stage peer review and physical GPU lab validation.

Includes Free AI Starter Kit

The Weekly AI Engineering Briefing

Join AI engineers building with Claude, MCP, Gemini, and open-source models. Received by developers, researchers, and technical founders.