Back to work

Open agent infrastructure · registry and CLI

Agent Mag Skills Ecosystem

A public distribution system for 492 installable AI-agent capabilities, built around declarative skill bundles, catalog APIs, CLI installation, validation, documentation, usage signals, and automated release workflows.

Founder and product engineerPublic ecosystem · active development

The registry count is a reviewed repository snapshot. Skills are declarative configuration bundles; installing one does not silently grant credentials or execute arbitrary code.

architectureSkill bundle → validation and sync → catalog API → CLI install → local agent

492

public installable skill bundles

9

CLI command modules

16

product sync, release, and review workflows

4

standalone public builder tools

The system

From operational problem to reliable workflow.

Problem

Agent capabilities are often trapped in private prompts, copied manually between tools, or distributed without a consistent contract. Builders need a way to inspect, install, update, and contribute skills without turning a prompt file into an opaque executable package.

Approach

Agent Mag treats a skill as a declarative configuration bundle: metadata, instructions, tool requirements, optional configuration, and documentation. GitHub is the reviewable source layer; validation and sync workflows feed the product catalog; APIs and the Agent Mag CLI provide discovery and installation into supported local agents.

Outcome

The ecosystem now exposes 492 skill bundles through a public repository and product catalog, backed by nine CLI command modules, automated sync and release paths, and a separate tools collection for reusable developer workflows.

Architecture

System architecture

The Skills Ecosystem is the reusable capability layer of Agent Mag. A skill packages operational instructions and tool expectations into a format that can be reviewed in source control, presented in a product catalog, installed with a CLI, and adapted to a supported agent runtime. The architecture keeps community distribution separate from credential custody and runtime authority.

Declarative skill bundle

Defines the skill identity, operating instructions, tool requirements, configuration contract, and human-readable usage notes in inspectable files.

manifest.json · prompt.md · tools.json · config.json · README.md

Source and review layer

Stores community skills in a public repository where changes can be diffed, reviewed, versioned, and discussed before distribution.

GitHub · pull requests · MIT-licensed registry

Validation and synchronization

Checks bundle shape and metadata, rejects malformed entries, and synchronizes accepted skills into Agent Mag's product-facing catalog.

Node.js scripts · schema checks · GitHub Actions

Catalog and API layer

Supports search, categories, detail views, rankings, installation metadata, and programmatic access from the web product and CLI.

Next.js · REST routes · Cosmos DB · search metadata

Installation client

Resolves a requested skill, previews its files, and writes the bundle into a supported local agent configuration without owning the user's provider credentials.

Agent Mag npm CLI · nine command modules

Adoption signal

Reports bounded installation events so maintainers can distinguish useful capabilities from registry inventory alone.

Install API · catalog counters · product analytics

Execution model

End-to-end execution

  1. 01

    Author the capability contract

    A contributor describes what the skill does, the instructions it gives an agent, the tools it expects, and any user-supplied configuration.

  2. 02

    Review the bundle in source

    The public repository makes prompt, manifest, and tool changes visible before they enter the distribution path.

  3. 03

    Validate and synchronize

    Automation checks required fields and normalizes accepted metadata before updating the Agent Mag catalog and documentation surfaces.

  4. 04

    Discover through web or CLI

    A builder searches by capability, reads the contract and requirements, and selects a named version rather than copying an untracked prompt.

  5. 05

    Install into the local runtime

    The CLI resolves the bundle and places configuration files into the supported agent environment; credentials stay under the builder's control.

  6. 06

    Measure and improve

    Bounded install signals and community changes inform ranking, documentation, compatibility work, and future product priorities.

Implementation

What Harshitha implemented

  • Structured a skill as a multi-file, declarative bundle so metadata, instructions, tools, optional configuration, and documentation can evolve independently.
  • Connected the public skills repository to product routes that support discovery, detail pages, categories, leaderboards, and install tracking.
  • Developed the product contract for CLI commands that add skills, tools, documentation context, and local model instructions.
  • Built automation around synchronization, review, publishing, indexing, and generated documentation instead of relying on manual catalog updates.
  • Kept runtime secrets outside skill bundles and documented the distinction between installing configuration and authorizing external services.
  • Extended the ecosystem beyond skills through public tools, roadmap proposals, documentation, and community-health repositories under the same organization.

Contribution summary

  • Defined the product direction connecting an open GitHub registry, Agent Mag product surfaces, API routes, documentation, and CLI distribution.
  • Structured skills as inspectable configuration rather than arbitrary executable code, preserving a clear boundary between instructions, tool declarations, and the user's runtime.
  • Built catalog and installation journeys so builders can evaluate a skill before adding it to their local agent environment.
  • Established automated validation, synchronization, documentation, and release workflows around a community-contribution model.
  • Connected registry usage signals to the wider Agent Mag platform so installs can become product feedback rather than anonymous file copies.

Failure design

Reliability engineering

R01

Risk

A malformed community bundle breaks installation or catalog rendering.

Control

Required manifest fields, normalized metadata, and validation in the sync path reject incomplete bundle shapes before distribution.

R02

Risk

Installing a skill is confused with granting it authority.

Control

Bundles remain declarative and exclude credentials; the user's agent runtime and provider permissions continue to control execution.

R03

Risk

Web catalog data drifts from the source repository.

Control

Automated synchronization and generated documentation reduce manual duplication between GitHub, APIs, and product surfaces.

R04

Risk

A bundle update silently changes behavior.

Control

Source-controlled files, pull-request diffs, and explicit package metadata make behavior changes inspectable and reviewable.

R05

Risk

Registry size is mistaken for real adoption.

Control

Installation events are tracked separately from inventory so ranking and roadmap decisions can use actual usage signals.

Tradeoffs

Technical decisions

Use declarative bundles instead of executable packages

Prompts and tool contracts should be inspectable before installation; execution authority belongs to the user's runtime, not the registry artifact.

Keep GitHub as the contribution source

Pull requests provide an understood review, history, and attribution model while the Agent Mag product can focus on discovery and installation.

Offer both web and CLI discovery

The web product supports evaluation and comparison; the CLI shortens the path from a selected capability to a configured local agent.

Separate skills from tools

Instruction bundles and executable developer utilities have different security and installation expectations and should not share an ambiguous package type.

Credibility

Evidence and scope

Attribution boundary

Harshitha owns the Agent Mag ecosystem's product direction and organization. The 492-skill figure describes the public registry snapshot, including community-ready configuration bundles; it is not a claim that she manually authored every individual skill file.

Technology

TypeScriptNext.jsNode.jsnpm CLIGitHub ActionsMDXJSON manifestsREST APIsOpenClawClaude CodeCodex