convez.ai
MCP ServerDevOps & InfraestruturaAvançado
quzhiii

MCPskills Center

quzhiii/mcpskills-center

Governança local de MCP servers e agent skills com scan, audit e sincronização segura

Ver no GitHub
InicialPluginScore · de 0 a 100

Plano de controle local para gerenciar MCP servers e skills de agentes. Escaneia, audita, planeja e sincroniza capacidades distribuídas em múltiplas ferramentas (Claude Code, OpenCode, CodeBuddy, etc.) através de CLI com segurança por padrão.

Instalação

$npx convez-ai quzhiii/mcpskills-center

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
2
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 4 meses
Versão
v0.2.3
Issues abertas
0
Contribuidores
1

Casos de uso

  • Inventariar e auditar MCP servers em múltiplos agentes AI
  • Detectar e consolidar skills duplicadas com segurança
  • Planejar e sincronizar distribuição de capabilities sem risco
  • Gerar relatórios de saúde e conformidade offline

Veredito

Ferramenta robusta e operacional para quem gerencia múltiplos agentes AI em produção. Prioriza segurança (read-only por padrão, backups timestamped) e auditoria, mas requer familiaridade com CLI e estrutura de MCP servers.

README

A tradução para o português deste README ainda está na fila.

MCPskills Center

Local-first CLI for scanning, auditing, planning, and synchronizing agent skills while inventorying MCP servers across Claude Code, OpenCode, Codex, CodeBuddy, WorkBuddy, Trae, Qoder, and Qoder Work.

Runtime Language Mode Default Outputs

中文文档 · English

Quickstart · Outputs · Commands · Scenarios · Supported Agents · Profiles · Safety · Boundaries


What Is This?

MCPskills Center gives one local machine a clear control surface for agent capabilities that are usually scattered across multiple tools.

It scans installed MCP servers and skill directories, normalizes their metadata, highlights duplicates and broken entries, generates dry-run skill sync plans, checks health status, and renders readable reports for review.

The product direction is CLI-first. The CLI remains the governance kernel and operational source of truth for scan, plan, apply, and restore. A local Web console can wrap these artifacts later, but it should not replace the CLI execution model.

Claude Code / OpenCode / Codex ─┐
CodeBuddy / WorkBuddy / Trae ──┼─→ scan → audit → plan → verify → report
Qoder / Qoder Work ────────────┘                      │
                                                      ├─→ sync dry-run
                                                      ├─→ sync apply with backup manifest
                                                      ├─→ restore from manifest
                                                      └─→ offline dashboard.html

The current release focuses on a practical local workflow:

CapabilityWhat it gives you
Inventory scanningUnified view of MCP servers, skills, install paths, metadata, and issues
Audit reportingDuplicate skills, duplicate MCPs, missing SKILL.md, symlink review items, sensitive env key risk
Sync planningCanonical skill distribution plan with dry-run output before any writes
Safe apply / restoreExplicit --confirm, approved-root checks, timestamped backups, restore manifests
ProfilesRead-only planning for scenario-based capability bundles such as coding or research
Health checksPassive validation by default, explicit active command probing when allowlisted
DashboardStatic offline HTML report at reports/dashboard.html

Agent support status is summarized in docs/supported-agents.md.


Governance Roadmap

Current write-capable priority remains skills governance: make duplicate skill installs explainable, reversible, and safe to consolidate through sync --dry-run, sync --apply --confirm, and sync --restore.

MCP governance is active on the read-only/report-first lane. The read-only MCP kernel is complete: scope-aware governance, canonical profile evidence, write-readiness evidence, and deterministic canonical target policy are all on master. The MCP write model design (types, adapter interface, safety contracts) is also complete. The next MCP milestone is mcp-write-apply-v1: runtime implementation of MCP config apply/restore with per-adapter serialization for write-ready agents (Claude Code, OpenCode, Codex).

Longer-term layers are a local Web control plane over the CLI kernel, then intelligent local agent routing after governed capability state is stable. Web, SQLite history, and routing are intentionally not the first implementation priority.


Quickstart

Clone the repository, install dependencies, run the test suite, then generate the first inventory snapshot.

git clone https://github.com/quzhiii/mcpskills-center.git
cd mcpskills-center

npm install
npm test
npm run scan

Expected result:

  • TypeScript builds successfully into dist/.
  • Tests pass.
  • Generated reports appear under reports/.
  • reports/dashboard.html opens locally without external assets.

If you want a quick read-only sync plan next:

node dist/index.js sync --dry-run

Outputs

Main report set

scan writes:

  • reports/inventory-current.json
  • reports/inventory-current.md
  • reports/audit-current.md
  • reports/dashboard.html

sync --dry-run writes:

  • reports/sync-plan-current.json
  • reports/sync-plan-current.md

matrix writes:

  • reports/capability-matrix-current.json
  • reports/capability-matrix-current.md

sync --apply --confirm writes timestamped backup content under backups/ and a consolidated manifest for the apply run.

Dashboard preview

The static dashboard summarizes the current state in one offline HTML file:

Summary cards     Recommendations table     Skills table     Issues table
     │                     │                    │                │
     └───────────── all generated from the current local inventory ─────────────┘

Generated HTML is a reading surface. The JSON and Markdown reports remain the auditable artifacts for automation and review.


Commands

CommandPurposeWrites
npm run scanScan inventory, normalize records, audit findings, generate reportsreports/
npm run auditPrint audit summary in the terminalNone
node dist/index.js sync --dry-runGenerate a sync plan without changing agent configreports/
node dist/index.js sync --apply --confirmApply the current sync plan with backupsbackups/
node dist/index.js sync --restore <manifest>Restore a previous apply run from its manifestExisting targets
node dist/index.js profile listList available local profilesNone
node dist/index.js profile show <name>Print one profile JSONNone
node dist/index.js profile plan <name>Compare a profile against the current inventoryNone
node dist/index.js agents listList registered local agents from config/agents.jsonNone
node dist/index.js agents discoverDiscover local agent config candidates such as Qoder, CodeBuddy, WorkBuddy, and Traereports/
node dist/index.js matrixBuild a capability matrix across registered agents for discovered skills and MCP serversreports/
node dist/index.js healthRun passive MCP health checksNone
node dist/index.js health --active --allow-command <cmd> [--timeout <ms>]Run explicit active command probes for allowlisted commandsNone
node dist/index.js route <task>Recommend which agent to use for a taskNone
node dist/index.js helpShow CLI helpNone

Scenarios

1. Inspect the current machine state

npm run scan

Use this when you want the baseline inventory, audit report, and offline dashboard in one pass.

2. Review duplicates and risky entries in the terminal

npm run audit

Use this when you want a quick summary of duplicate skills, duplicate MCPs, missing SKILL.md, symlink review items, and sensitive env key risk.

3. Plan a canonical skill sync without writing anything

node dist/index.js sync --dry-run

Optional custom canonical directory:

node dist/index.js sync --dry-run --canonical-dir C:\path\to\canonical-skills

The current planner uses a canonical store plus distribution actions to agent skill roots. Review the generated sync plan before any apply step.

If you later run sync --apply with a custom canonical directory, add that directory to config/sync.json under approvedSyncRoots first. Apply validates both source and target paths.

4. Apply the current sync plan with backups

node dist/index.js sync --apply --confirm

Apply mode recomputes the current plan, backs up existing targets when present, and writes one manifest that can be used for restore.

5. Restore a previous apply run

node dist/index.js sync --restore C:\path\to\manifest.json

Restore mode copies backed-up content from the manifest back to the original target path after approved-root validation.

6. Plan a profile for a focused workflow

node dist/index.js profile plan coding

This reports already-present, missing, and disable actions for the named profile without changing any live config.

7. List registered local agents

node dist/index.js agents list

Use this to review the currently loaded registry without scanning live config. The checked-in config/agents.json also contains disabled read-only entries for Qoder, Qoder Work, CodeBuddy, WorkBuddy, and Trae, but disabled entries are filtered out of the runtime-loaded list.

8. Discover local agent candidates

node dist/index.js agents discover

Discovery is read-only. It checks known local paths and writes reports/agent-discovery-current.json plus reports/agent-discovery-current.md.

9. Build a cross-agent capability matrix

node dist/index.js matrix

Use this when you want a cross-agent capability view for the current inventory. It writes reports/capability-matrix-current.json and reports/capability-matrix-current.md with per-agent presence, missing coverage, and shared-capability counts.

10. Run safe passive MCP health checks

node dist/index.js health

Passive mode checks transport shape, command presence, URL validity, and sensitive env key risk without spawning commands.

11. Run explicit active command probes

node dist/index.js health --active --allow-command npx --timeout 3000

Active mode only probes allowlisted commands with --version, uses argument arrays instead of shell strings, and times out when the command does not return.


Profiles

Sample profiles live in config/profiles/ and are evaluated read-only against the current inventory.

ProfilePurposeAgents
codingCore development workflow with testing and debugging supportclaude-code, opencode, codex
researchInvestigation and web-reading workflowclaude-code, opencode
lark-officeFeishu/Lark and document-production workflowclaude-code
securityDefensive review and security-audit workflowclaude-code

Example:

node dist/index.js profile list
node dist/index.js profile show coding
node dist/index.js profile plan coding

Sync Approval Config

Writable sync roots are controlled by config/sync.json.

{
  "approvedSyncRoots": [
    "config/canonical-skills",
    "C:/Users/quzhi/.claude/skills",
    "C:/Users/quzhi/.opencode/skills",
    "C:/Users/quzhi/.codex/skills"
  ]
}

Relative paths are resolved from the project root. Invalid values fail before an apply run begins.


Safety Model

  • Default behavior is read-only or dry-run.
  • sync --apply requires explicit --confirm.
  • sync --restore requires an explicit manifest path.
  • Apply and restore only operate inside approved roots.
  • Existing targets are backed up before overwrite when a previous target exists.
  • Generated reports never print secret values.
  • Sensitive env handling only reports key-risk presence.
  • Passive health checks do not spawn commands.
  • Active health checks require --active.
  • Active probes only succeed for commands explicitly allowlisted with --allow-command.
  • --timeout is optional and defaults to 3000 ms when omitted.
  • Active health probes run spawn(command, ['--version'], { shell: false }) semantics rather than shell strings.

Repository Layout

mcpskills-center/
├── config/
│   ├── profiles/
│   └── sync.json
├── docs/
├── fixtures/
├── reports/
├── backups/
├── src/
├── README.md
└── README.zh-CN.md

Fixture policy:

  • .claude/ is treated as local machine configuration and is not tracked as live repository payload.
  • Repository-tracked skill samples live under fixtures/skills/.
  • Those fixtures are synthetic and intentionally minimal.

Boundaries

  • The current sample config/sync.json is machine-oriented and may need adjustment on another system.
  • Profile matching supports short MCP ids such as playwright against project-scoped ids such as C:/Users/quzhi:playwright.
  • Passive HTTP and SSE health checks validate preserved URL or host values when present in config.
  • Active health checks validate command reachability through --version; they do not perform a full MCP handshake.
  • OpenCode array-form commands are normalized to the leading executable name for health probing.
  • There is no packaged binary yet; use node dist/index.js ... after build.

Documentation

DocumentPurpose
docs/MCPskills-center-background-and-plan.mdProduct background, machine context, and original project framing
docs/migration-notes.mdMigration decisions and retained / excluded assets
docs/plans/2026-06-03-mcpskills-center-completion.mdImplementation plan for the completed CLI workflow

Verify Locally

npm run build
npm test
npm audit --audit-level=moderate

For a workflow smoke check after the test suite:

npm run scan
npm run audit
node dist/index.js sync --dry-run
node dist/index.js health

License

MIT

Ver o original no GitHub

Bom para

Relacionados