TokRepo — Summary

TokRepo is an open registry for AI assets (skills, prompts, MCP configs, scripts, workflows) designed for agent-to-tool and agent-to-agent discovery, where AI agents can find, inspect, and install reusable capabilities using typed contracts, with no need for manual browsing.

Problem it solves: AI agents have no standard way to discover what reusable capabilities (skills, prompts, MCP configs) exist outside their current session — TokRepo provides a machine-readable registry with multiple discovery surfaces (CLI, MCP server, well-known URLs, planning-time fallback) so agents can autonomously find and install capabilities during task planning.

Distinctive trait: The only registry in the batch designed for agent-initiated discovery: agents call tokrepo agent-check "<task>" --json during planning to find relevant capabilities, and tokrepo agent-handoff --json after tasks to return improved capabilities back to the registry.

Target audience: AI agents (not just humans) performing autonomous task planning, plus developers who want a universal asset registry compatible with Claude Code, Codex, Cursor, Gemini CLI, Cline, Windsurf, and 10+ other coding agents.

Scope: This GitHub repo is a landing page pointing to three separate repos: CLI (tokrepo-cli), MCP server (tokrepo-mcp-server), and cross-platform skill (tokrepo-search-skill). The registry website is tokrepo.com.

Differs from seeds: Uniquely positioned as infrastructure rather than a methodology framework. Unlike all seeds (which provide skills/methodology), TokRepo is the registry layer that other skills frameworks would use for asset distribution — closer in architecture to npm than to superpowers or BMAD-METHOD.

TokRepo — Overview

Origin

Created by henu-wang (single contributor). Stars: 0 (GitHub landing page repo only). The actual product has 3 repos, a website (tokrepo.com), npm packages, and an official MCP Registry listing (io.github.henu-wang/tokrepo-mcp-server). Active as of 2026-05-18.

Philosophy

From the README:

"TokRepo is designed to become infrastructure for agent-to-tool and agent-to-agent work: humans give the task, agents plan autonomously, discover reusable capabilities through TokRepo, install only through typed contracts, and hand reusable improvements back after confirmation."

Agent Discovery Design

Multiple machine-readable surfaces:

• tokrepo.com/.well-known/tokrepo.json — Manifest
• tokrepo.com/.well-known/agent.json — Portable agent manifest
• tokrepo.com/.well-known/agent-card.json — A2A agent card
• tokrepo.com/.well-known/mcp/server.json — MCP server manifest
• tokrepo.com/mcp.json — Root MCP manifest
• tokrepo.com/.well-known/tool-catalog.json — Tool catalog
• tokrepo.com/agents.txt — Agent text entry
• tokrepo.com/llms.txt — LLM crawler entry

Bootstrap Pattern

npx -y tokrepo@latest init-agent --target all

Writes managed instructions for 12+ platforms: Codex, Claude Code, Gemini CLI, Cursor, GitHub Copilot, Cline, Windsurf, Roo, OpenHands, Aider, MCP clients, and generic agents.

Why This GitHub Repo Exists

From the README:

"Most GitHub discovery still happens through repository names, README text, topics, and linked repos. This repo is the canonical GitHub landing page for TokRepo so that developers can understand the product quickly."

The repo itself is a discovery artifact — it exists to be found by GitHub search, not to ship code.

TokRepo — Architecture

Component Repos

This GitHub repo (henu-wang/tokrepo) is a landing page. The actual architecture spans three separate repos:

1. CLI (tokrepo-cli)

• npm: npm install -g tokrepo or npx tokrepo
• Commands: search, install, agent-check, agent-handoff, init-agent
• Connects to tokrepo.com API

2. MCP Server (tokrepo-mcp-server)

• npm: npx tokrepo-mcp-server
• Official MCP Registry: io.github.henu-wang/tokrepo-mcp-server
• Add to agents:

  claude mcp add tokrepo -- npx tokrepo-mcp-server
  codex --mcp-server tokrepo -- npx tokrepo-mcp-server
  gemini settings mcp add tokrepo -- npx tokrepo-mcp-server
  

3. Cross-Platform Skill (tokrepo-search-skill)

• Variants for: Claude Code, Codex, Cursor, Gemini CLI, AGENTS.md
• Enables agents without MCP to use TokRepo via CLI fallback

Discovery Surfaces

tokrepo.com/
├── .well-known/
│   ├── tokrepo.json          # Primary manifest
│   ├── agent.json            # Cross-agent manifest
│   ├── agent-card.json       # A2A agent card
│   ├── mcp/server.json       # MCP server description
│   └── tool-catalog.json     # Full tool catalog
├── mcp.json                  # Root MCP manifest
├── agents.txt                # Agent discovery
├── llms.txt                  # LLM crawler
└── agent-instructions/
    └── tokrepo.md            # Agent instructions

Bootstrap Flow

npx -y tokrepo@latest init-agent --target all

Writes agent instruction files for all supported platforms (12+) so future agents know to use TokRepo during planning.

This Landing Page Repo Structure

henu-wang/tokrepo/
├── .github/
├── AGENTS.md       # Cross-agent instructions (points to primary targets)
├── LICENSE         # MIT
└── README.md       # Primary discovery surface

Required Runtime

• node>=18 / npm for CLI and MCP server
• Network access to tokrepo.com

TokRepo — Components

CLI Commands (tokrepo-cli)

MCP Server Tools (tokrepo-mcp-server)

Available tools via MCP when server is added:

• Search registry
• Install assets
• (exact tool list not visible without fetching the separate repo)

Cross-Platform Skill Variants

The tokrepo-search-skill repo provides skill files for:

• Claude Code
• Codex
• Cursor
• Gemini CLI
• AGENTS.md (generic)

Asset Types in Registry

• Skills
• Prompts
• MCP configs
• Scripts
• Workflows

This Repo's Components

Zero skills, commands, hooks, or MCP servers bundled in this repo directly.

TokRepo — Uniqueness

Differs from Seeds

TokRepo occupies a category entirely separate from all 11 seeds: it is registry infrastructure, not a methodology or skill framework.

• Unlike all seeds (superpowers, spec-kit, BMAD, claude-flow, etc.) — which provide skills/commands/agents — TokRepo provides the discovery and distribution layer for those skills.
• Closest seed in spirit: taskmaster-ai (Archetype 3 — MCP-anchored toolserver) because both use MCP as a primary interface. But taskmaster-ai does task decomposition; TokRepo does asset registry.
• The agent-check + agent-handoff loop is unique in the corpus: no seed framework has an agent-to-registry feedback cycle.

Positioning

TokRepo is the only framework in the batch that explicitly designs for:

1. Agent-as-consumer — agents discover and install capabilities during task planning
2. Agent-as-contributor — agents submit improvements back to the registry after task completion
3. Cross-platform bootstrap — init-agent --target all configures 12+ agent platforms in one command

This is "npm for agents" with an additional bidirectional loop.

Observable Failure Modes

1. Zero stars, 0 contributors — this specific landing page repo has no traction, though the product (tokrepo.com) may have independent adoption
2. Registry dependency — all discovery requires network access to tokrepo.com; no offline mode
3. Thin GitHub surface — the repo itself ships nothing; developers who find it via GitHub search see only a pointer to other repos
4. Agent-handoff UX — the "human confirmation required" step in agent-handoff may break autonomous agent workflows

Explicit Antipatterns

None stated.

Most Interesting Design Choice

The .well-known/agent-card.json support — implementing the A2A (Agent-to-Agent) protocol for registering TokRepo itself as a discoverable agent capability. This is forward-looking infrastructure assuming a future where agents discover each other via A2A endpoints.