gptadmin_opensource MCP server
Commit Nov 11, 2026A self-hosted MCP hub that enables AI assistants to manage infrastructure, run code, and execute server administration tasks.
Also known as megamen32/gptadmin_opensource, gptadmin-opensource
- Stars
- 11
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 11, 2026
- License
- AGPL-3.0
What is gptadmin_opensource MCP?
gptadmin_opensource is a self-hosted MCP hub that enables AI assistants to manage infrastructure, run code, and execute server administration tasks.
Summary generated from the project's README
About gptadmin_opensource MCP server
GPT-Админ is a self-hosted MCP hub designed to bridge AI assistants with your physical or virtual infrastructure. It acts as a central command center where you can plug in various MCP tools like shellmcp, chrome-devtools, or openmemory to give your AI agent direct control over servers running Linux, macOS, or Windows. Instead of just generating commands for you to copy, this hub allows the agent to read system state, execute commands, and validate the actual output.
Developers can connect their preferred AI models through three distinct adapters. You can use the OpenAI Action for ChatGPT and Open WebUI, the MCP remote SSE adapter for Claude, Codex, or OpenCode, or a browser extension to connect free web-based models like DeepSeek, Qwen, or GigaChat. The setup is managed locally, ensuring your tokens and server data remain within your own infrastructure. Authentication is handled via Bearer tokens or OAuth flows to maintain security during administration tasks.
What you can do with it
- 1Automated Server AdministrationAsk an AI to manage system services or network configurations directly on your host.
- 2Incident Diagnosis and RepairInstruct an agent to find the cause of a service failure by reading system logs.
- 3Code Execution and DebuggingHave an AI assistant write, run, and validate code or launch subagents to fix bugs.
- 4Web-based Research and DocumentationUse an AI to browse the web and read documentation to solve technical problems.
Written from the project's README. Every use case is checked against a quote from it.
Automated checks
Ran Nov 14, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachablemegamen32/gptadmin_opensource is public and reachablegithub.com/megamen32/gptadmin_opensource ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)AGPL-3.0spdx.org/licenses/AGPL-3.0.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-11-11github.com/megamen32/gptadmin_opensource ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubGPT‑Админ
One MCP hub — any AI controls any infrastructure.
GPT‑Админ is a self-hosted MCP hub that sits between your AI assistant and your servers. Plug servers and any MCP tools into the hub, then connect your AI via one of three adapters. Manage everything from server admin to running subagents — from ChatGPT, Claude, Codex, DeepSeek, Qwen, even Yandex Alice or Sber GigaChat.
🌐 Website & live docs: https://became.bezrabotnyi.com 📚 Full docs: docs/ — architecture, adapters, hub, install, config, API, security, FAQ 📦 Install:
curl -s https://raw.githubusercontent.com/megamen32/gptadmin_opensource/main/deploy/install.sh | bash
Why
Most "AI server admin" tools are either cloud-only, tied to one AI, or require a paid API. GPT‑Админ is different:
- Self-hosted — your servers, your hub, your tokens. Nothing leaves your infra.
- Any AI — ChatGPT, Claude, Codex, OpenCode, DeepSeek, Qwen, Alice, GigaChat. Even free web chats work via a browser extension.
- MCP-native — the hub speaks MCP remote SSE, so any MCP client connects. Plug in chrome-devtools, openmemory, or any other MCP too.
- Real execution — not "here's the command, copy it". The agent reads state, runs commands, validates, and reports actual output.
- Optional relay and webhook capabilities — when you need them, GPTAdmin can expose a single MCP server or route webhook events through dedicated documented surfaces. Start from the quick path below; see the linked docs only if you need those extra capabilities.
How it works
MCP tools plug IN AIs connect OUT (3 adapters)
┌─────────────────┐ ┌──────────────────────┐
│ shellmcp │ │ Claude · Codex │ (MCP client)
│ chrome-devtools │ ──► ┌──┴──────────────┐ │
│ openmemory │ │ GPT‑Админ │ ──► │ DeepSeek · Qwen │ (browser ext)
│ any MCP │ ◄── │ MCP hub │ │ Alice · GigaChat │
└─────────────────┘ └──┬──────────────┘ ──► │ ChatGPT · OpenUI │ (OpenAI Action)
│ │ └──────────────────────┘
▼
your servers
(Linux · macOS · Windows)
The hub is one process. Tools plug into it. AIs connect to it via one of three adapters. Capabilities (admin, code, logs, search, subagents) come from the hub — independent of which AI you use.
Quickstart
# 1. Install (Linux / macOS — auto-detects user/system mode)
curl -s https://raw.githubusercontent.com/megamen32/gptadmin_opensource/main/deploy/install.sh | bash
# Windows (PowerShell, no Administrator needed)
iwr -UseBasicParsing https://became.bezrabotnyi.com/install_win.ps1 | iex
The installer prints one Hub URL. Open that URL, then pick the shortest path for the client you want:
ChatGPT Custom GPT → import https://your-hub.example/actions/openapi.yaml
That Actions file is generated from the Hub's current MCP tool list. It is the schema the GPT should import; you do not hand-write it.
For authentication, choose one of these two paths:
Bearer → paste only a scoped token value issued by the Hub
OAuth → use the Hub authorize/token flow from the Hub URL
Canonical MCP/OAuth connector identity
For the shared ChatGPT/Codex/native-MCP connector, the canonical public origin
is https://mcp.bezrabotnyi.com and the MCP endpoint is:
https://mcp.bezrabotnyi.com/mcp
Use that same origin when following OAuth discovery and reconnecting a client.
The per-tunnel u-f<ID>.t.became.bezrabotnyi.com hostname is a legacy FRP
transport/recovery fallback, not the connector identity for new ChatGPT or
Codex setup. Do not replace the canonical origin with a loopback Hub URL or a
transient tunnel name.
If an existing connector still points at an u-f<ID> URL, update it to the
canonical endpoint above and complete OAuth again. Verify the live metadata at
https://mcp.bezrabotnyi.com/.well-known/oauth-authorization-server before
retrying the connector.
An existing legacy compatibility token remains valid until its owner explicitly rotates or removes it.
Never paste an internal service secret into a GPT or client. If you need a different adapter, use the dedicated docs below.
# 2. Connect your AI — pick one adapter:
# - OpenAI Action (ChatGPT Custom GPT / Open WebUI) → see docs
# - MCP remote SSE (Claude / Codex / OpenCode) → see docs
# - Browser extension (free web AIs) → see docs
# 3. Use it:
# "поставь nginx", "почини сайт", "покажи логи", "запусти codex для фикса бага"
Full per-adapter instructions: https://became.bezrabotnyi.com/#/docs
Администрирование и CloudOS
Основная веб-консоль — https://your-hub.example/admin/. Она защищена
настроенным паролем администратора и включает настройки, серверы, доступ и
операции Hub. Корневой адрес Hub направляет туда же.
CloudOS не входит в обычный hub-пакет: это отдельный аддон примерно на 56 MiB.
Если desktop-интерфейс нужен, соберите tools/build_cloudos_ui.sh, распакуйте
build/gptadmin-cloudos-addon.tar.gz рядом с Hub и запустите ./install.sh.
После установки он доступен авторизованному администратору по /cloudos/.
What you can do through the hub
| Capability | Example |
|---|---|
| Server administration | restart systemd services, manage firewall/nginx/fail2ban/sshd, install packages |
| Write & run code | edit files, run type-check/lint/build, launch subagents ("run codex to fix this bug") |
| Code Mode | chain authorized MCP calls in a short JavaScript script and return only the selected JSON result |
| Profile context | deliver current instructions before execution; reuse context with compact results |
| Fix & clean PRs | find fork in memory, keep one feature commit, force-push via SSH |
| Check logs | parse journalctl/nginx/postgres, find anomalies, suggest & apply fixes |
| Web search | via chrome-devtools MCP — agent opens pages, reads docs, searches |
| Diagnose incidents | find downed service, read logs, understand cause (ECONNREFUSED, 503, OOM), fix it |
Three adapters
| Adapter | For | How |
|---|---|---|
| OpenAI Action | ChatGPT, Open WebUI | Import the generated OpenAPI schema, then choose Bearer or OAuth. No Codex limits. |
| MCP remote SSE | Claude Desktop, Codex, OpenCode | Hub is an MCP server (Streamable HTTP). Add to claude_desktop_config.json. |
| Browser extension | DeepSeek, Qwen, Alice, GigaChat, ChatGPT (free) | Userscript (Tampermonkey/Firefox) adds MCP buttons to web chat UIs. No paid API. |
All three connect to the same hub. Same capabilities. Pick what fits your AI.
Install paths
| OS | user-mode (default) | system-mode (sudo) |
|---|---|---|
| Linux / macOS | ~/.local/share/gptadmin |
/opt/gptadmin |
| Windows | %LOCALAPPDATA%\gptadmin |
C:\Program Files\gptadmin |
The installer auto-detects the mode: no sudo → user-mode (home dir, user
service); sudo → system-mode. No domain needed — auto-tunnel via FRP or
Cloudflare gives a public URL.
Project layout
go-hub/ # the Go MCP hub — proxies commands to agents
server_for_installer.py # serves install scripts + OpenAPI
gptadmin_security.py # auth, OAuth, token validation
cli.py # `gptadmin` CLI (setup, tunnel, status, logs)
telegram_logs_bot.py # optional Telegram alerts
go-shellmcp/ # primary shell agent (Go) — runs on target machines
public/ # OpenAPI schema, install scripts, mcp-bridge.user.js
deploy/ # install scripts, systemd, nginx configs
tests/ # test_hub.py, test_rootd.py
docs/ # OPEN_CORE_PLAN.md
Structure is being cleaned up — see
docs/OPEN_CORE_PLAN.mdfor the target layout and the open-core launch plan.
Running from source
python3 -m venv .venv && source .venv/bin/activate
pip install fastapi uvicorn requests
# hub (terminal 1)
python3 cli.py setup --hub --tunnel none --user
# agent on a target machine (terminal 2)
python3 cli.py setup --no-hub --shellmcp --hub-url http://127.0.0.1:9001 --user
# smoke test (terminal 3)
python tests/test_hub.py
Contributing
PRs welcome. See CONTRIBUTING.md for dev setup, code style, and the PR process. For security issues see SECURITY.md.
License
GNU AGPL-3.0 — see LICENSE.
The hub, shellmcp, all three adapters, and the basic web panel are free forever. Future paid offerings (hosted cloud, enterprise SSO/audit/RBAC, advanced panel, support) will live in a separate repo — the core stays open.
Links
- 🌐 Website & docs: https://became.bezrabotnyi.com
- 💬 Telegram: @careviolan
- 📦 Other projects: https://bezrabotnyi.com
Failover: alive and degraded
GPTAdmin can keep a fallback control plane available when the primary hub dies. A fallback node watches the public health URL, promotes only after a failure threshold, and keeps enough of the service online to inspect logs, reach surviving servers, and restore the primary.
This is graceful degradation, not magic multi-master replication: recent in-memory counters or jobs running exactly on the dead node can be stale. But queues, background job metadata, large stdout/stderr spill files, shellmcp spool files, outbox responses, registry snapshots and failover logs are written to disk, so the recovery trail is not lost forever.
See docs/FAILOVER.md.
Show current public URLs
Use the CLI to print the active hub URL, tunnel mode, MCP endpoints and per-server OpenAPI Action schema URLs:
sudo gptadmin urls
sudo gptadmin urls --all
sudo gptadmin urls --json
The default output focuses on the hub and ShellMCP endpoints. --all also includes every registered MCP server such as OpenMemory, FileShare and Chrome DevTools.
Optional capabilities
If you later need a single-server relay or webhook ingress, use the dedicated docs:
gptadmin_opensource MCP questions
What AI models can I connect to the hub?
You can connect ChatGPT, Claude, Codex, DeepSeek, Qwen, Yandex Alice, and Sber GigaChat.
Is my data safe with this tool?
Yes, it is self-hosted, meaning your servers, hub, and tokens remain within your own infrastructure.
What kind of server administration tasks can it perform?
It can restart systemd services, manage firewalls, nginx, fail2ban, and sshd, as well as install packages.