pendnt MCP server
2026년 9월 15일 커밋헤드리스 Claude Code 에이전트를 위한 지속 가능한 사용자 승인, 알림 및 이벤트 기반 웨이크업 기능을 제공합니다.
다른 이름 Pendnt/pendnt-plugin, pendnt-plugin
- 스타
- 0
- Tools
- —
- 다운로드, 30일
- —
- 최신 버전
- 0.2.0
- 마지막 커밋
- 2026년 9월 15일
- 라이선스
- MIT
pendnt MCP란?
pendnt은(는) MCP server입니다. 헤드리스 Claude Code 에이전트를 위한 지속 가능한 사용자 승인, 알림 및 이벤트 기반 웨이크업 기능을 제공합니다.
프로젝트 README에서 생성된 요약
pendnt MCP server 정보
pendnt는 Claude Code 세션을 원격 워크스페이스에 연결하여 헤드리스 에이전트를 관리할 수 있는 방법을 제공합니다. 터미널 없이 권한 프롬프트를 처리해야 하는 대화형 또는 헤드리스 Claude Code 인스턴스 실행 개발자를 위해 설계되었습니다. 이 MCP server를 사용하면 에이전트가 지속 가능한 시스템을 통해 사용자 승인을 요청하고, 알림을 보내며, 이벤트 기반 웨이크업을 트리거할 수 있습니다.
사용자는 도구 사용 전 자동으로 실행되는 PermissionRequest hook을 구현하거나 전용 permission_prompt MCP 도구를 사용하여 이를 적용할 수 있습니다. 설정에는 PENDNT_API_KEY가 필요하며, PENDNT_URL을 통해 프로덕션 API 또는 로컬 개발 인스턴스를 가리키도록 구성할 수 있습니다. 이 서버에는 request_approval, notify, schedule_wakeup, kv_set 등 14개의 도구가 포함되어 있어 모델이 워크스페이스 상태와 통신을 직접 제어할 수 있습니다.
이 솔루션은 장시간 실행되는 승인 프로세스를 지원한다는 점에서 독특합니다. 표준 도구 호출은 타임아웃이 발생할 수 있지만, permission_prompt 도구는 호출을 최대 540초 동안 유지할 수 있습니다. 또한 fail-closed 보안 모델을 따르므로, 네트워크 오류, 키 누락 또는 타임아웃이 발생하면 요청을 거부하여 헤드리스 에이전트가 승인되지 않은 작업을 수행하지 않도록 보장합니다.
활용 사례
- 1헤드리스 도구 실행 승인개발자는 모바일 기기나 웹 인박스에서 요청을 승인함으로써 헤드리스 Claude Code 에이전트가 명령을 계속 수행하도록 허용할 수 있습니다.
- 2에이전트 상태 업데이트 수신시스템이 알림이나 실행 완료 메시지를 외부 엔드포인트로 전달하여 에이전트의 진행 상황을 실시간으로 확인할 수 있습니다.
프로젝트의 README를 바탕으로 작성되었습니다. 모든 활용 사례는 해당 내용의 인용문과 대조하여 확인합니다.
pendnt MCP 설치 방법
형식 확인: 2026년 11월 10일Cursor에 추가Goose에 추가Claude Code에 pendnt MCP 설치
$ claude mcp add --transport http pendnt-plugin https://api.pendnt.dev/mcpCursor에 pendnt MCP 설치
~/.cursor/mcp.json
{
"mcpServers": {
"pendnt-plugin": {
"url": "https://api.pendnt.dev/mcp"
}
}
}VS Code에 pendnt MCP 설치
.vscode/mcp.json
{
"servers": {
"pendnt-plugin": {
"type": "http",
"url": "https://api.pendnt.dev/mcp"
}
}
}VS Code CLI에 pendnt MCP 설치
$ code --add-mcp '{"name":"pendnt-plugin","type":"http","url":"https://api.pendnt.dev/mcp"}'Codex에 pendnt MCP 설치
$ codex mcp add pendnt-plugin --url https://api.pendnt.dev/mcpCodex config.toml에 pendnt MCP 설치
~/.codex/config.toml
[mcp_servers.pendnt-plugin] url = "https://api.pendnt.dev/mcp"
Gemini CLI에 pendnt MCP 설치
$ gemini mcp add -s user --transport http pendnt-plugin https://api.pendnt.dev/mcpGemini settings.json에 pendnt MCP 설치
~/.gemini/settings.json
{
"mcpServers": {
"pendnt-plugin": {
"url": "https://api.pendnt.dev/mcp",
"type": "http"
}
}
}구성
MCP_TOOL_TIMEOUTPENDNT_API_KEYPENDNT_URL
| 이름 | 설명 | 유형 |
|---|---|---|
| MCP_TOOL_TIMEOUT | — | |
| PENDNT_API_KEY | — | |
| PENDNT_URL | — |
자동 검사
2026년 11월 15일에 실행됨전체 보고서 →아직 인증되지 않았습니다. 서버는 80점 이상의 점수, 심각한 결함 없음, 그리고 검사에 충분한 데이터를 갖추어야 합니다. 검사 방식 · 점수를 높이기 위해 수정할 항목
- 소스 저장소 연결 및 접근 가능pendnt/pendnt-plugin is public and reachablegithub.com/pendnt/pendnt-plugin ↗통과
- 패키지가 이 저장소를 소스로 명시검사되지 않음
- 빌드 출처(Sigstore 증명)검사되지 않음
- 오픈 소스 라이선스 (OSI)MITspdx.org/licenses/MIT.html ↗통과
- 최근 1년 내 커밋 또는 릴리스Last commit or release 2026-09-15github.com/pendnt/pendnt-plugin ↗통과
- OpenSSF Scorecard검사되지 않음
- 패키지 이름이 인기 패키지와 유사하지 않음검사되지 않음
- 패키지 및 해당 종속성에 중요한 보안 권고 없음검사되지 않음
- 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 응답HTTPS endpoint answered 401 (sign-in required)통과
README
GitHub에서 보기pendnt (Claude Code plugin)
Try it in 10 seconds — no account:
curl https://api.pendnt.dev/try?plain=1returns a 72-hour trial key (20 requests) with a quickstart.
Wires a Claude Code session — interactive or headless (claude -p) — up
to a pendnt workspace: durable human approvals, notifications, and a "run finished" ping,
without the agent needing a terminal.
"pendnt" is a placeholder product name. It's kept in exactly one place — PRODUCT_NAME in
lib/config.mjs — plus the name fields in
plugin.json, marketplace.json and
.mcp.json, and the matching MCP server name registered in
/app/src/routes/mcp.ts. Rename all of those together and it's renamed
everywhere that matters.
This plugin ships two independent ways to answer permission prompts headlessly — pick one (or both;
--permission-prompt-tool takes priority when both are wired up, see §1a below):
- A
PermissionRequesthook (hooks/permission-request.mjs) — works with plainclaude -p, no extra flags. Runs automatically before every tool-use prompt. - The
permission_promptMCP tool, via--permission-prompt-tool mcp__pendnt__permission_prompt— Claude Code's own dedicated mechanism for headless permission prompts (seeINTEGRATIONS.md§1a).
Plus a Notification hook and a Stop hook that both relay to POST /v1/notify — see
What's included below.
Install
Local marketplace (for trying this out before publishing anywhere)
This repo doubles as its own local marketplace — .claude-plugin/marketplace.json
lists this plugin with source: "./", pointing at the same directory plugin.json lives in. From
Claude Code:
/plugin marketplace add Pendnt/pendnt-plugin
/plugin install pendnt@pendnt
(or non-interactively: claude plugin marketplace add Pendnt/pendnt-plugin && claude plugin install pendnt@pendnt)
To publish for real later, split marketplace.json out into its own repo with source pointing at
this plugin's repo (see INTEGRATIONS.md §1d for the marketplace
manifest shape and github/git-subdir/npm/archive source types), then
claude plugin marketplace add <owner>/<repo> from anywhere.
Env vars
| Variable | Required | Default | Meaning |
|---|---|---|---|
PENDNT_API_KEY |
yes | — | Authorization: Bearer <key> for every call to the API (aio_..., from POST /dev/bootstrap locally, or your real workspace's API key in production). Without it, the PermissionRequest/Notification/Stop hooks all fail closed (deny / no-op) rather than erroring the session. |
PENDNT_URL |
no | https://api.pendnt.dev |
Base URL of the pendnt REST API — not including /mcp or /v1. Point this at http://localhost:8787 for local dev against wrangler dev (see below). |
PENDNT_APPROVAL_TIMEOUT_S |
no | 600 (10 min) |
Total wall-clock budget the PermissionRequest hook spends re-polling before it gives up and denies. Each individual poll is capped server-side at 25s (wait_s), so the hook loops GET /v1/requests/:id?wait_s=25 until this budget runs out. |
Set them in your shell, or in .claude/settings.json under env — see
settings.json docs. .mcp.json's ${PENDNT_URL:-...} /
${PENDNT_API_KEY} syntax is Claude Code's own env-var expansion for .mcp.json values (headers
and URLs) — see INTEGRATIONS.md §1e.
What's included
.mcp.json— registers thependntremote MCP server (${PENDNT_URL:-https://api.pendnt.dev}/mcp,Authorization: Bearer ${PENDNT_API_KEY}) so all 14 tools (request_approval,check_request,permission_prompt,notify,create_endpoint,list_endpoints,wait_for_event,schedule_wakeup,list_wakeups,cancel_wakeup,kv_get,kv_set,kv_delete,whoami) are available to the model directly, undermcp__pendnt__<tool>— exceptpermission_prompt, which Claude Code removes from its own visible tool list once named by--permission-prompt-tool(see below), since it's meant to be called by the CLI itself, not by the model.hooks/hooks.json— wires three hooks (shape matchessettings.json'shookskey, seeINTEGRATIONS.md§1d):PermissionRequest→hooks/permission-request.mjs. Fires before every tool-use prompt. POSTs a summary of the tool call toPOST /v1/requests(kind: "approval",wait_s: 25,timeout_s: $PENDNT_APPROVAL_TIMEOUT_S), then re-pollsGET /v1/requests/:id?wait_s=25until it's answered or the totalPENDNT_APPROVAL_TIMEOUT_Sbudget elapses. Prints exactly one JSON object perINTEGRATIONS.md§1b'sPermissionRequestcontract:
or{ "hookSpecificOutput": { "hookEventName": "PermissionRequest", "decision": { "behavior": "allow", "updatedInput": {} } } }{"behavior": "deny", "message": "..."}. Fails closed on every error path (missing API key, network failure, operator denies, request expires, timeout) — the only way to getallowis an explicit approval. An operator using the answer page's free-text box (always shown, even forkind: "approval"requests) is treated as approval only if the text starts with yes/y/allow/approve/ok — anything else denies with that text as the reason.Notification→hooks/notify.mjs. Relaysnotification_type/messagetoPOST /v1/notify. Cannot block Claude Code (per the hook's own semantics) and never tries to — always exits 0.Stop→hooks/stop.mjs. Relays a"run finished"message toPOST /v1/notifyat the end of every turn. Never returns{"decision":"block"}— it only notifies, it doesn't keep the agent going.
lib/— the three hook scripts' shared bits:config.mjs(env vars + thePRODUCT_NAMEconstant),client.mjs(thinfetchwrappers for/v1/requestsand/v1/notify),stdin.mjs(hook stdin JSON parsing). No npm dependencies — plain Node 18+ (fetch,AbortController) is enough.
The --permission-prompt-tool path (§1a)
Instead of (or in addition to — Claude Code prefers --permission-prompt-tool over hooks when both
apply) the PermissionRequest hook, you can point Claude Code's dedicated permission-prompt mechanism
directly at the permission_prompt MCP tool this plugin's .mcp.json registers — the one tool in this
server purpose-built for this flag (see /app/README.md's permission_prompt section for the full
contract):
export PENDNT_API_KEY=aio_...
export PENDNT_URL=https://api.pendnt.dev # or http://localhost:8787 for local dev
export MCP_TOOL_TIMEOUT=600000 # ms — must exceed PERMISSION_PROMPT_WAIT_S (default 540s)
claude -p --permission-prompt-tool mcp__pendnt__permission_prompt "deploy to prod"
Three things worth knowing before you rely on this path:
- Claude Code blocks the first turn until the
pendntMCP server connects, capped byMCP_TIMEOUT(default 30s, env var, milliseconds) — bump it if your network is slow:export MCP_TIMEOUT=60000(fromINTEGRATIONS.md§1a). - As of CLI ≥2.1.199, an
allowfrom a--permission-prompt-toolis converted todenyfor any MCP tool call flagged_meta["anthropic/requiresUserInteraction"]— this prompt tool can't rubber-stamp those regardless of what the operator answers (also fromINTEGRATIONS.md§1a). permission_promptholds the MCP tool call open server-side for up toPERMISSION_PROMPT_WAIT_S(env var on/app, default 540s = 9 min) waiting for the operator, unlike every other tool's ≤25swait_s. Claude Code's ownMCP_TOOL_TIMEOUT(its per-tool-call timeout, ms, default 30000) must be set higher than that or the CLI gives up first — henceexport MCP_TOOL_TIMEOUT=600000above.
Unlike request_approval, permission_prompt does return Claude Code's exact {behavior, updatedInput, updatedPermissions, interrupt} / {behavior:"deny", message, interrupt} contract natively — an allow
echoes the original input back as updatedInput, a deny (or an unanswered timeout) carries a
human-readable message. That's the same translation the PermissionRequest hook below does by hand for
its own hook-shaped contract; permission_prompt does it as the MCP tool result directly, which is what
--permission-prompt-tool actually reads.
Headless example
export PENDNT_API_KEY=aio_...
export PENDNT_URL=https://api.pendnt.dev
export MCP_TOOL_TIMEOUT=600000
claude -p --permission-prompt-tool mcp__pendnt__permission_prompt \
"review open PRs and merge the ones that pass CI"
Or, with the PermissionRequest hook doing the work instead (no extra flag needed once the plugin is
installed):
export PENDNT_API_KEY=aio_...
export PENDNT_URL=https://api.pendnt.dev
claude -p "review open PRs and merge the ones that pass CI"
Either way, every tool-use prompt now creates a durable approval request instead of blocking on a
terminal that doesn't exist — check your configured channels (email/web inbox — see /app/README.md's
"Channels" section for what's actually implemented today), answer from your phone, and the agent
resumes.
Running the hook locally against wrangler dev
From /app:
cd ../app
npm install
npm run db:migrate:local
npm run dev # http://localhost:8787
In another terminal, bootstrap a workspace + API key (see /app/README.md) and point the plugin's env
vars at it:
BOOT=$(curl -s -X POST http://localhost:8787/dev/bootstrap -d '{}')
export PENDNT_API_KEY=$(echo "$BOOT" | jq -r .api_key)
export PENDNT_URL=http://localhost:8787
Then either run a hook script directly, feeding it the JSON a real Claude Code hook invocation would send on stdin (useful for iterating on the hook without a full Claude Code session):
echo '{"session_id":"s1","cwd":"'"$PWD"'","tool_name":"Bash","tool_use_id":"toolu_1","tool_input":{"command":"deploy prod"}}' \
| PENDNT_APPROVAL_TIMEOUT_S=60 node hooks/permission-request.mjs
...or install the plugin locally (/plugin marketplace add Pendnt/pendnt-plugin from Claude Code, see
Install) and drive a real session — with PENDNT_URL/PENDNT_API_KEY exported in the
shell Claude Code runs in, every tool-use prompt round-trips through your local wrangler dev
instance. Answer the request the same way /app/README.md's curl walkthrough does: compute the HMAC
signature yourself (it's a pure function of the request id and SIGNING_SECRET) and POST /a/<id>/<sig>, or watch the log delivery channel's console output from wrangler dev for the
answer link once a real channel is wired up.
Caveats
- No npm dependencies by design (
fetch/AbortControlleronly) — every hook script runs under plainnode hooks/*.mjs, nonpm installstep inside the plugin itself. - The
PermissionRequesthook's fail-closed defaults (missing key, network error, timeout → deny) are a deliberate safety choice, not a limitation to work around — a headless agent with nobody to ask should not proceed by default. hooks/hooks.json'sNotificationmatcher is not narrowed to specificnotification_types — every notification (including ones that aren't permission-prompt-related) gets relayed. Narrow the matcher or filter insidehooks/notify.mjsif that's noisier than you want.
Guides
Practical writeups on running agents unattended, at pendnt.dev/guides: --dangerously-skip-permissions alternatives · claude -p on cron · the --permission-prompt-tool contract · hooks for unattended runs · exit codes: retry vs stay dead
pendnt MCP 질문
pendnt MCP를 Claude Code에 추가하려면 어떻게 하나요?
터미널에서 claude mcp add --transport http pendnt-plugin https://api.pendnt.dev/mcp를 실행합니다. 모든 프로젝트에서 사용할 수 있게 하려면 --scope user를 추가합니다.
pendnt MCP를 Cursor에 추가하려면 어떻게 하나요?
~/.cursor/mcp.json를 열고 mcpServers 아래에 서버를 추가합니다. 위 설치 섹션에 표시된 Cursor 구성을 사용합니다.
계정 없이 pendnt를 체험해 볼 수 있나요?
특정 curl 명령어를 실행하면 20회 요청이 가능한 72시간 체험 키를 받을 수 있습니다.
API 키가 없거나 네트워크 오류가 발생하면 어떻게 되나요?
시스템은 fail-closed 방식으로 설계되었습니다. 즉, 에이전트가 작업을 계속 진행하거나 세션 오류를 일으키는 대신 요청을 거부하거나 아무 작업도 수행하지 않습니다.