TokRepo MCP Server

> Agent-native MCP server for TokRepo: session bootstrap, capability discovery, search/detail, trust verification, install planning, Codex staging, lifecycle inspection/update/uninstall/rollback, handoff/harvest planning, and human-confirmed publishing of reusable AI assets.

Quick Start

Claude Code

claude mcp add tokrepo -- npx -y tokrepo-mcp-server

Cursor / Windsurf

Add to your MCP config (`~/.cursor/mcp.json`):

{
  "mcpServers": {
    "tokrepo": {
      "command": "npx",
      "args": ["-y", "tokrepo-mcp-server"]
    }
  }
}

OpenAI Codex / Gemini CLI

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

What It Does

Once connected, your AI assistant can:

• Discover during planning — turn a task or capability gap into structured candidate skills, prompts, MCP configs, scripts, and workflows
• Resolve capability gaps — select a candidate with verification evidence, install plan, lifecycle contract, next MCP calls, and CLI fallbacks before local build
• Search public AI assets by keyword or category with agent fit signals
• Browse trending assets, filter by type (MCP, Skill, Prompt, Agent, Script)
• Get details — full documentation, install instructions, and metadata
• Verify trust — read-only content hash, install plan hash, permission envelope, policy, trust_score_v2, `evidence_bundle`, SBOM-lite, `signature_evidence`, blockers, and warnings
• Plan before install — get install plan v2 with policy decisions, rollback, verification, `evidence_bundle`, SBOM-lite, and `signature_evidence`
• Safe Codex install — dry-run by default; risky assets must be staged or explicitly approved
• Lifecycle control — list, update, uninstall, and roll back managed Codex installs
• Post-task harvest — inspect changed or explicit local files with `tokrepo_harvest` before any push, including private package drafts, `quality_gate`, package manifest, SBOM-lite, and provenance
• Project memory — pair with `tokrepo init-agent --target all` so future agents know to call TokRepo during planning
• Discovery surfaces — official MCP `server.json`, A2A agent card, portable agent manifest, tool catalog, `.well-known`, `agents.txt`, `llms.txt`, and npm metadata are kept machine-readable for agents and registries
• Funnel visibility — anonymous aggregate events show whether agents actually discover, plan, install, hand off, and push reusable assets

Available Tools

Example Conversations

You: "Find me a good MCP server for databases"
AI: [calls tokrepo_resolve_capability] → Ranks DBHub, Supabase MCP, PostgreSQL MCP and returns verification/install-plan evidence for the selected candidate

You: "What video assets should I install?"
AI: [calls tokrepo_resolve_capability] → Finds relevant skills, checks fit, trust, and policy, then asks before installing

You: "Install that cursor rules asset"
AI: [calls tokrepo_verify] → Checks trust_score_v2, permissions, blockers, and warnings
AI: [calls tokrepo_install_plan] → Reviews policy and actions
AI: [calls tokrepo_codex_install with dry_run=false, confirm=true] → Writes only after explicit confirmation
AI: [calls tokrepo_rollback with dry_run=true] → Shows exactly what would be removed before rollback

You: "We created a reusable project rule; save it for future agents"
AI: [calls tokrepo_harvest] → Returns explicit files, hashes, quality_gate, package drafts, SBOM-lite, provenance, metadata defaults, and private-by-default push guidance
AI: [asks for confirmation before tokrepo_push] → Uploads only reviewed files

Make Future Agents Discover TokRepo

Run this once in a project:

npx tokrepo init-agent --target all

It writes managed instructions to `AGENTS.md`, `CLAUDE.md`, `GEMINI.md`, Cursor rules, GitHub Copilot instructions, Cline rules, Windsurf rules, Roo rules, OpenHands microagents, Aider conventions, `.mcp.json`, and `.tokrepo/agent.json`. The rule is simple: during planning, when the agent sees a capability gap, it should call `tokrepo_resolve_capability` or `tokrepo_discover` before inventing a one-off local tool. After a task, agents can call `tokrepo_harvest` or run `tokrepo harvest --changed --json` to suggest reusable files for user-confirmed private publishing.

Machine-Readable Listing

Registries and agents can discover this server through:

• MCP server manifest: tokrepo.com/.well-known/mcp/server.json
• TokRepo manifest: tokrepo.com/.well-known/tokrepo.json
• Portable agent manifest: tokrepo.com/.well-known/agent.json
• A2A agent card: tokrepo.com/.well-known/agent-card.json
• Tool catalog: tokrepo.com/.well-known/tool-catalog.json
• Trust manifest: tokrepo.com/.well-known/tokrepo-trust.json
• Default agent policy pack: tokrepo.com/policy-packs/default-agent-policy.json
• Eval evidence: tokrepo.com/evals/agent-discovery.json
• Multi-agent compatibility: tokrepo.com/evals/multi-agent-compatibility.json
• Agent memory schema: tokrepo.com/schemas/agent-memory.schema.json
• Agent evidence bundle schema: tokrepo.com/schemas/agent-evidence-bundle.schema.json
• Capability resolution schema: tokrepo.com/schemas/capability-resolution.schema.json
• Harvest report schema: tokrepo.com/schemas/harvest-report.schema.json
• Handoff package schema: tokrepo.com/schemas/handoff-package.schema.json
• Agent text entry: tokrepo.com/agents.txt
• Agent instructions: tokrepo.com/agent-instructions/tokrepo.md
• Agent ecosystem distribution pack: tokrepo.com/agent-ecosystem.json
• LLM crawler entry: tokrepo.com/llms.txt

Use `https://tokrepo.com/agent-ecosystem.json` for agent marketplace submissions, starter templates, README snippets, install guides, and example projects. It contains canonical listing copy, ecosystem channels, target project-memory files, and verification commands.

TokRepo emits anonymous aggregate funnel events for `tokrepo_resolve_capability`, `tokrepo_discover`, `tokrepo_verify`, `tokrepo_install_plan`, install dry-runs, installs, harvests, handoffs, and pushes. It does not send task text or file contents. Disable with `TOKREPO_TELEMETRY=0`.

Why TokRepo?

TokRepo is the open registry for AI assets — like npm for packages, but for AI skills, prompts, MCP configs, and workflows.

• Curated assets — quality-reviewed, not a dump
• Agent-native — every asset has machine-readable install contracts
• Universal — works with Claude Code, Cursor, Codex, Gemini CLI, and any MCP client
• CLI available — `npx tokrepo search "query"` / `npx tokrepo install `

Requirements

• Node.js >= 18
• Internet connection (queries tokrepo.com API)

Links

• Website: tokrepo.com
• CLI: npm: tokrepo
• GitHub: henu-wang/tokrepo-mcp-server
• API: tokrepo.com/.well-known/tokrepo.json

License

MIT