설치 신뢰도 평가, 공급망 기록 검증, stdio 트래픽 공격 패턴 검사를 통해 MCP server를 보호합니다.
다른 이름 getmcpm/cli, cli, @getmcpm/cli
- 스타
- 3
- Tools
- —
- 다운로드, 30일
- 5,354
- 최신 버전
- 0.54.0
- 마지막 커밋
- 2026년 11월 14일
- 라이선스
- MIT
mcpm MCP란?
mcpm은(는) MCP server입니다. 설치 신뢰도 평가, 공급망 기록 검증, stdio 트래픽 공격 패턴 검사를 통해 MCP server를 보호합니다.
프로젝트 README에서 생성된 요약
mcpm MCP server 정보
mcpm은 AI 클라이언트를 사용하는 개발자가 설치 신뢰도를 평가하고 트래픽을 검사하여 Model Context Protocol server를 안전하게 보호할 수 있도록 지원합니다. 이 CLI 도구는 보안 가드 및 패키지 매니저 역할을 하며, 사용자가 server를 검색하고 Sigstore를 통해 공급망 기록을 검증하며 도구 정의 드리프트(drift)를 탐지할 수 있게 합니다. 계정, 모델 구독 또는 텔레메트리 없이 로컬에서 결정론적으로 작동하지만, 패키지 확인을 위해 공개 MCP registry 및 npm에 접속합니다.
사용자는 Claude Desktop, Cursor, VS Code, Windsurf를 포함한 6가지 클라이언트 어댑터의 설정을 관리할 수 있습니다. 이 도구는 stdio 기반 server의 트래픽을 래핑하여 공격 패턴을 식별함으로써 런타임 보호를 제공합니다. macOS 사용자의 경우, 선택 사항인 confinement 프로필을 통해 프로세스 제한을 추가하여 server 기능을 더욱 제한할 수 있습니다. 개발자는 mcpm install과 같은 명령어로 클라이언트 설정을 작성하고, mcpm guard enable로 런타임 검사 및 피닝(pinning)을 활성화합니다.
활용 사례
- 1설치 전 server 보안 평가개발자는 후보 MCP server를 조사하여 신뢰 점수와 수행되지 않은 보안 검사 항목을 파악할 수 있습니다.
- 2도구 정의 변경으로부터 보호시스템은 악성 업데이트를 방지하기 위해 런타임 중 도구 스키마나 어노테이션의 예기치 않은 변경을 모니터링합니다.
- 3멀티 클라이언트 설정 관리사용자는 서로 다른 AI 개발 환경 전반에 걸쳐 MCP server 정의의 설정 및 동기화를 자동화할 수 있습니다.
- 4macOS에서 server 권한 제한개발자는 설치된 server에 제한된 프로세스 제한을 적용하여 특정 비밀 경로에 대한 접근을 차단할 수 있습니다.
프로젝트의 README를 바탕으로 작성되었습니다. 모든 활용 사례는 해당 내용의 인용문과 대조하여 확인합니다.
mcpm MCP 설치 방법
형식 확인: 2026년 11월 10일Cursor에 추가Goose에 추가Claude Code에 mcpm MCP 설치
$ claude mcp add cli -- npx -y @getmcpm/cliCursor에 mcpm MCP 설치
~/.cursor/mcp.json
{
"mcpServers": {
"cli": {
"command": "npx",
"args": [
"-y",
"@getmcpm/cli"
]
}
}
}VS Code에 mcpm MCP 설치
.vscode/mcp.json
{
"servers": {
"cli": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"@getmcpm/cli"
]
}
}
}VS Code CLI에 mcpm MCP 설치
$ code --add-mcp '{"name":"cli","type":"stdio","command":"npx","args":["-y","@getmcpm/cli"]}'Claude Desktop에 mcpm MCP 설치
claude_desktop_config.json
{
"mcpServers": {
"cli": {
"command": "npx",
"args": [
"-y",
"@getmcpm/cli"
]
}
}
}Codex에 mcpm MCP 설치
$ codex mcp add cli -- npx -y @getmcpm/cliCodex config.toml에 mcpm MCP 설치
~/.codex/config.toml
[mcp_servers.cli] command = "npx" args = ["-y", "@getmcpm/cli"]
Gemini CLI에 mcpm MCP 설치
$ gemini mcp add -s user cli npx -- -y @getmcpm/cliGemini settings.json에 mcpm MCP 설치
~/.gemini/settings.json
{
"mcpServers": {
"cli": {
"command": "npx",
"args": [
"-y",
"@getmcpm/cli"
]
}
}
}자동 검사
2026년 11월 14일에 실행됨전체 보고서 →자동 검사를 통과했습니다. 이는 보안 감사가 아닙니다. 코드를 직접 읽고 서버에 필요한 최소한의 권한만 부여하십시오. 검사 방식
- 소스 저장소 연결 및 접근 가능getmcpm/cli is public and reachablegithub.com/getmcpm/cli ↗통과
- 패키지가 이 저장소를 소스로 명시@getmcpm/cli names getmcpm/cli as sourcenpmjs.com/package/@getmcpm/cli ↗통과
- 빌드 출처(Sigstore 증명)Sigstore attestation for @getmcpm/clinpmjs.com/package/@getmcpm/cli/v/0.54.0#provenance ↗통과
- 오픈 소스 라이선스 (OSI)MITspdx.org/licenses/MIT.html ↗통과
- 최근 1년 내 커밋 또는 릴리스Last commit or release 2026-11-13github.com/getmcpm/cli ↗통과
- OpenSSF ScorecardOpenSSF Scorecard 7.8 / 10; weakest: Code-Review 0, CII-Best-Practices 0, Vulnerabilities 3scorecard.dev/viewer/?uri=github.com/getmcpm/cli ↗통과
- 패키지 이름이 인기 패키지와 유사하지 않음No popular package one edit away from @getmcpm/cli통과
- 패키지 및 해당 종속성에 중요한 보안 권고 없음the package checked; no critical advisoryosv.dev ↗통과
- 모델에 대한 숨겨진 지침 없음0 tool descriptions and the README read; no override phrases, credential paths or hidden characters통과
- 파괴적인 도구에 라벨 표시됨검사되지 않음
- 원격 엔드포인트가 MCP 핸드셰이크에 응답검사되지 않음
README
GitHub에서 보기mcpm — MCP security guard and package manager
Assess MCP servers before installing, verify supply-chain records, and inspect wrapped stdio traffic for known attack patterns on your machine. mcpm combines client configuration, trust evidence, tool-definition drift detection and an opt-in macOS confinement profile in one CLI.
The default assessment and guard path is deterministic and local: no account, model subscription, mcpm backend or telemetry. Registry resolution and package record checks contact the public MCP registry and npm; server launchers may fetch packages independently. An external scanner runs only if you explicitly name an already-installed executable.
Visit the project homepage, or preview and deploy it locally.
Quick start
Requires Node ^22.22.2 || ^24.15.0 || >=26.0.0: Node 22.22.2+, 24.15.0+
or 26+. Other versions, including 23 and 25, are outside the supported range.
npm install -g @getmcpm/cli
mcpm search filesystem
mcpm info io.github.domdomegg/filesystem-mcp
mcpm why io.github.domdomegg/filesystem-mcp
mcpm install io.github.domdomegg/filesystem-mcp
Search shows registry lifecycle status; info and why explain computed trust.
Review the findings and checks that did not run before accepting an installation.
The install command writes client configuration; it does not automatically enroll
the server in runtime protection. Registry names and available versions can change.
mcpm guard enable --dry-run
mcpm guard enable
# Restart the affected MCP server in your client to load the wrapper.
mcpm guard status
Use --client <name> or --server <name> to narrow enrollment. Status reports
recognized configuration and pin state; it does not prove your client loaded
that configuration. New servers start with first-session pinning pending.
The relay supports stdio only. HTTP/SSE servers have no runtime inspection;
explicit --allow-unguarded consent records that absence, without adding
protection. Warnings forward the original traffic; blocks substitute an error.
Optional guard enable --confine adds bounded macOS process restrictions;
network-enabled launchers can still send accessible data, including configured
credentials. See guard coverage and confinement.
The binary is mcpm; the package is @getmcpm/cli. The Homebrew formula
mcpm belongs to a different project. Use this scoped npm package or the
pnpm, npx and mise alternatives.
What it does
| Task | Evidence or behavior |
|---|---|
| Assess before installation | Metadata patterns for secrets, injection and suspicious package/schema shapes, with score components and unrun checks |
| Lock and verify | Package-version records, integrity and applicable Sigstore build-identity baselines; client-free CI verification |
| Guard MCP traffic | Explicit stdio wrapping, field-scoped patterns, first-observation tool pins, drift checks and policy-controlled actions |
| Manage client configuration | Detect, install, import, enable/disable and compare definitions across six client adapters |
| Inventory project exposure | Read-only scope source/instruction metadata: Claude Code default, explicit VS Code/Copilot selector |
| Review local artifact changes | Experimental archive-only skill/plugin snapshot comparison with byte/declaration coverage and unknown host state |
These operations answer different questions. An install-time metadata assessment does not execute the server or audit its source. A complete inventory does not establish active selection. A matching lock does not establish safe behavior.
Security boundaries and evidence
A trust score is a signal about available evidence, not a safety probability. Search listing is not publisher verification. Current assessment commands do not run health checks; an unavailable external scanner does not earn a clean pass. See score math, scanner protocol and safety floors.
Tool pins use trust on first observation (TOFU) at runtime, not approval at install. Initial malicious definitions can become the baseline. Field-level comparison warns on description-only changes and blocks schema/annotation changes; unannounced same-session changes and older coarse pins can also block. Pattern findings can independently block. Review the actual upgrade before accepting drift; do not automatically re-pin merely to clear a refusal.
The guard inspects supported JSON-RPC carriers, not every action a server takes. Unknown embedded request shapes produce coverage warnings; binary media, external schema references and fetched archives are outside complete inspection. Retrieved resource/prompt text has warn-and-forward limits. A response block cannot undo server-side effects that already happened. See the carrier matrix and signature catalog.
Confinement is opt-in and macOS-only. Its standard profile denies selected secret paths and restricts writes, with launcher-dependent network access. It is not general exfiltration prevention or an agent-wide sandbox. Backend absence can produce an explicit unconfined launch for non-required enrollment; required and integrity failures have separate refusal rules. See the precise confinement conditions.
Sigstore verification authenticates build identity, not safe code. Integrity checks compare published package records; mcpm does not intercept the bytes a package manager downloads at server launch. Unkeyed local integrity sidecars detect corruption but cannot authenticate against a same-user writer who can rewrite both files.
The public mcp-guardbench contains
versioned results and known misses measured through published guard inspect
binaries. Those stateless runtime results do not measure install-time detection
prevalence or promise coverage of unseen attacks. The
v0.53.0 release record retains the
published artifact, compatibility and synchronized companion evidence.
Common workflows
| Goal | Start here |
|---|---|
| Explain one candidate | mcpm why <name> |
| Review installed configuration and findings | mcpm list, mcpm doctor, mcpm audit (coverage report: --report-json) |
| Find available updates | mcpm outdated |
| Compare clients | mcpm sync --check |
| Start a project stack | mcpm init, edit mcpm.yaml, then mcpm lock |
| Install a reviewed stack with frozen checks | mcpm up --frozen |
| Check a committed lock in CI | mcpm verify --json |
| Inspect captured MCP frames without a server | mcpm guard inspect frames.ndjson --json |
| Inventory a project | mcpm scope /path/to/project --json |
| Inventory fixed VS Code/Copilot sources | mcpm scope /path/to/project --client vscode --json |
| Compare local skill snapshots | mcpm review-artifacts before.tar after.tar --profile agent-skills-2026-11-06 --json |
init scaffolds an empty stack; it does not install curated starter packs. To
share an existing setup, mcpm export > mcpm.yaml, review the declaration, then
lock it. Locks distinguish MCP publication versions from actual package versions;
OCI tags remain mutable. See stack examples and the command index
and use mcpm <command> --help for flags.
For CI, use the verification Action.
Pin its source and executed CLI version separately. Verification checks lock
coverage as well as applicable integrity/provenance baselines; unknown or
unenforceable coverage must remain visible. audit --sarif provides a separate
report for GitHub code scanning. Exit codes and JSON are documented in
contracts.
When a guard block fires, read its signature and remediation. A legitimate upgrade
may need explicit accept-drift with the reported hash; a false positive may need
a scoped policy override. Blocks/warnings are recorded in
~/.mcpm/guard-events.jsonl. Pause and mute changes are startup snapshots:
restart the affected server to load changes or resume after a pause expires.
See guard commands and policy.
Supported clients
| Adapter | Managed configuration scope |
|---|---|
| Claude Desktop | Native user configuration |
| Claude Code | User-global mcpServers in ~/.claude.json |
| Cursor | User-global ~/.cursor/mcp.json |
| VS Code | User MCP configuration, including JSONC comments/trailing commas |
| Windsurf | User-global MCP configuration |
| Gemini CLI | User-global ~/.gemini/settings.json |
macOS, Linux and Windows paths are supported by the adapters. Run mcpm doctor
for detected configurations. Portable/workspace and Copilot Agent Host paths are
not managed by those adapters; the read-only scope inventory has its own fixed
coverage. Unreadable configuration is not evidence of an empty or healthy setup.
Adapter support and native proof are separate. Pinned disposable-client runs cover named VS Code/Claude Code paths, with separate Cursor and Gemini evidence; they do not establish all clients/platforms or live-model behavior. Cursor/Gemini disable parks managed definitions for future discovery; it does not stop running peers or override project-local/native exclusion state. See native scope and platform limits, Cursor and Gemini.
Agent use
mcpm serve exposes search, info, install, list, remove, audit, doctor, setup and
up through stdio MCP. For clients using an mcpServers configuration map:
{
"mcpServers": {
"mcpm": {
"command": "npx",
"args": ["-y", "@getmcpm/cli@0.53.0", "serve"]
}
}
}
Other clients use their native schema. This pins the executed CLI; review updates explicitly. Restart your client after adding the configuration, then ask:
Find filesystem MCP servers. Explain findings, lifecycle status and checks that did not run. Show proposed configuration changes before installing.
Prefer versioned structured evidence, retaining unsupported schema versions,
unrun checks and incomplete coverage as unknown. A successful MCP transport is
not proof that every check passed. Served mutation tools retain trust floors;
CLI-only scope and review-artifacts are separate inspection surfaces.
See agent workflows and contracts.
The portable mcpm usage skill is self-contained. Copy its folder from a reviewed, pinned repository checkout into your host's skill directory. npm ships the CLI, not that skill. It adds no hooks, scripts or tool grants; discovery, permissions and model compliance depend on the host.
Agent lifecycle hooks cover a different boundary from MCP wire inspection. Command hooks can also be local and deterministic; some host hooks can deny or modify tool calls, while post-action hooks cannot undo completed effects. See relay, confinement and hook comparison for host coverage, failure limits and the bounded native experiments.
Documentation and contribution
Use the task-oriented docs index for security references, consumer guides, contribution/release workflows and dated evidence. Current selected work lives in BACKLOG; older roadmaps retain historical planning context.
For source setup and appropriate checks, read CONTRIBUTING. Contributor instructions/skills are repo-owned and separate from the distributed usage skill. Security findings should go through private vulnerability reporting; scanner false-positive/negative reports can use ordinary GitHub issues. See SECURITY for the reporting policy.
mcpm MCP 질문
mcpm MCP를 Claude Code에 추가하려면 어떻게 하나요?
터미널에서 claude mcp add cli -- npx -y @getmcpm/cli를 실행합니다. 모든 프로젝트에서 사용할 수 있게 하려면 --scope user를 추가합니다.
mcpm MCP를 Cursor에 추가하려면 어떻게 하나요?
~/.cursor/mcp.json를 열고 mcpServers 아래에 서버를 추가합니다. 위 설치 섹션에 표시된 Cursor 구성을 사용합니다.
mcpm은 무엇으로부터 보호하나요?
MCP server의 비밀 정보 노출, 인젝션, 의심스러운 패턴을 평가하고, 알려진 공격 패턴을 탐지하기 위해 stdio 트래픽을 검사합니다.
mcpm을 사용하려면 계정이나 텔레메트리가 필요한가요?
아니요. 기본 평가 및 guard 경로는 로컬에서 결정론적으로 실행되며 계정, 모델 구독, 백엔드 또는 텔레메트리를 요구하지 않습니다.