AI Skill Hub 推荐使用:开源MCP工具 是一款优质的MCP工具。AI 综合评分 7.5 分,在同类工具中表现稳健。如果你正在寻找可靠的MCP工具解决方案,这是一个值得深入了解的选择。
Enforced permissions for AI agents - policy + enforcement in one stack,提高AI安全性和可信赖性。
开源MCP工具 是一款遵循 MCP(Model Context Protocol)标准协议的 AI 工具扩展。通过 MCP 协议,它可以让 Claude、Cursor 等主流 AI 客户端直接访问和操作外部工具、数据源和服务,实现 AI 能力的无缝扩展。无论是文件操作、数据库查询还是 API 调用,都可以通过自然语言在 AI 对话中直接触发,极大提升生产效率。
Enforced permissions for AI agents - policy + enforcement in one stack,提高AI安全性和可信赖性。
开源MCP工具 是一款遵循 MCP(Model Context Protocol)标准协议的 AI 工具扩展。通过 MCP 协议,它可以让 Claude、Cursor 等主流 AI 客户端直接访问和操作外部工具、数据源和服务,实现 AI 能力的无缝扩展。无论是文件操作、数据库查询还是 API 调用,都可以通过自然语言在 AI 对话中直接触发,极大提升生产效率。
# 方式一:通过 Claude Code CLI 一键安装
claude skill install https://github.com/coproduct-opensource/nucleus
# 方式二:手动配置 claude_desktop_config.json
{
"mcpServers": {
"--mcp--": {
"command": "npx",
"args": ["-y", "nucleus"]
}
}
}
# 配置文件位置
# macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
# Windows: %APPDATA%/Claude/claude_desktop_config.json
# 安装后在 Claude 对话中直接使用 # 示例: 用户: 请帮我用 开源MCP工具 执行以下任务... Claude: [自动调用 开源MCP工具 MCP 工具处理请求] # 查看可用工具列表 # 在 Claude 中输入:"列出所有可用的 MCP 工具"
// claude_desktop_config.json 配置示例
{
"mcpServers": {
"__mcp__": {
"command": "npx",
"args": ["-y", "nucleus"],
"env": {
// "API_KEY": "your-api-key-here"
}
}
}
}
// 保存后重启 Claude Desktop 生效
| Tier | What | Isolation | Platform | Status |
|---|---|---|---|---|
| **0 — Scan** | nucleus audit [PATH] in CI | Static analysis of PodSpec / MCP / settings configs; CI exit codes | any | Usable today (92 tests) |
| **1 — Enforce** | nucleus run --local / agent PreToolUse hook | Local tool-proxy routes every call through the lattice; **process-level** env + command isolation (no VM) | any host | Alpha |
| **2 — Isolate** | nucleus run via nucleus-node | Firecracker microVM + network namespace + default-deny egress + seccomp | **Linux + KVM only** | Alpha |
Tier 1 is process-level (env-clear isolation,cap-std, command allowlist) — it is not kernel/container isolation and does not prevent kernel escapes or network exfiltration whenbashis allowed. The agent PreToolUse-hook variant of Tier 1 (nucleus run --hook,nucleus guard) is not runnable from this repo: it shells out to anucleus-claude-hookbinary that lives in the external private orchestrator, and the in-repo install hint (cargo install --path crates/nucleus-claude-hook) is stale — that crate does not exist here. Tier 2 isolation requires Linux with/dev/kvm; on macOS the node returns an explicit "firecracker requires Linux" error, and the macOS test suite exercises config/allocation logic, not a live VM boot.
---
cargo install --git https://github.com/coproduct-opensource/nucleus nucleus-cli
nucleus audit [PATH] # Tier 0: scan agent configs, no runtime (CI exit codes)
nucleus run --local "your task" # Tier 1: run with enforced permissions (process-level, no VM)
Every tool call flows through the permission kernel. nucleus run tracks data provenance and blocks dangerous combinations — like writing code derived from untrusted web content.
Vendor neutrality. The core library and the genericcredentials.env/PodSpecinterface are intended to be vendor-agnostic, but the surface is not fully clean today. Two things break it: (1) the reference agent runner shipped withnucleus run/shellis coupled to one specific assistant CLI (binary name, default model string, permission-bypass flag), and (2)nucleus-specstill hardcodes specific LLM-vendor hostnames and workload-identity defaults (see Known Gaps). A PreToolUse hook for AI coding assistants lives in the external private orchestrator nucleus-code, built on this runtime. Treat the runner as an integration example, not a vendor-agnostic component.
---
Every agent session can be packaged as a provenance bundle — the agent's payload plus a signed lineage envelope proving how it was produced. Each lineage edge (one per tool call, LLM call, or derived artifact) carries a child SPIFFE id encoding its derivation, an Ed25519 signature, and a prev_hash link; the whole log is committed to an RFC 9162 Merkle tree with signed tree heads, inclusion/consistency proofs, and optional external-witness cosignatures (Nucleus + C2SP tlog-witness).
The point is independent verification. verify_bundle re-checks the entire bundle — per-edge signatures, hash chain, session membership, Merkle inclusion, cosignature thresholds, and payload binding — against a trust anchor you supply out-of-band. The bundle's own embedded keys are deliberately ignored. The same audited Rust verifier ships three ways:
| Surface | What it is | Status |
|---|---|---|
nucleus-envelope (Rust) | Core verify_bundle + bundle builder | Working — 36 integration tests (73 total incl. unit tests) |
@coproduct/verify (WASM/JS) | verify(receipt, anchor) one-liner; runs in browser/Node with zero service trust | Builds + smoke-tested; **npm publish gated** |
nucleus-verifier (Python) | PyO3 backend binding | Builds; **no in-repo tests yet** |
nucleus-verifier-service (HTTP) | Optional convenience verifier + transparency log | Deploy-ready (fly.toml; 26 integration / 70 total tests); **not yet hosted** |
A CI-gated adversarial corpus of 8 forged bundles (tampered edges, swapped signatures, truncation, attacker JWKS, unknown kid, foreign parent, non-pod root) must be rejected on every merge — every_corpus_case_is_rejected passes today. This is the security promise, not a slogan. For replication, nucleus-bundle-cas fetches bundles by BLAKE3 root over bao-verified iroh-blobs QUIC so a peer can't substitute or truncate bytes (alpha; single-operator split-trust, no discovery/mesh, pins pre-1.0 iroh).
Honest scope. A green verification proves the lineage is authentic and intact. It does not prove the agent behaved well, that information-flow policy held, or that any computation was correct — those are separate guarantees. Issuing/signing identities is demo-only in this repo (dev-feature-gatedLocalIssuer); production needs an external SPIFFE issuer and witness. The public verifier service is not live (no hosted endpoint resolves today); it is self-hostable and deploy-ready. The@coproduct/verifynpm package and in-browser tamper demo are publish-gated; the compiled.wasmis a build product, not committed.
---
tlog-witness)nucleus工具提供了一个开源的MCP解决方案,提高了AI安全性和可信赖性,但其复杂性较高,需要有一定的Rust开发经验
AI Skill Hub 为第三方内容聚合平台,本页面信息基于公开数据整理,不对工具功能和质量作任何法律背书。
建议在沙箱或测试环境中充分验证后,再部署至生产环境,并做好必要的安全评估。
✅ MIT 协议 — 最宽松的开源协议之一,可自由商用、修改、分发,仅需保留版权声明。
总体来看,开源MCP工具 是一款质量良好的MCP工具,在同类工具中具备一定竞争力。AI Skill Hub 将持续追踪其更新动态,建议收藏备用,结合自身场景选择合适时机引入使用。
| 原始名称 | nucleus |
| 原始描述 | 开源MCP工具:Enforced permissions for AI agents - policy + enforcement in one stack。⭐16 · Rust |
| Topics | mcpagent-securityai-security |
| GitHub | https://github.com/coproduct-opensource/nucleus |
| License | MIT |
| 语言 | Rust |
收录时间:2026-06-05 · 更新时间:2026-06-05 · License:MIT · AI Skill Hub 不对第三方内容的准确性作法律背书。
选择 Agent 类型,复制安装指令后粘贴到对应客户端