devops-mcp

A mode-based MCP (Model Context Protocol) server that lets AI assistants

(Claude Desktop, Cursor, Windsurf, …) actually operate Linux servers

without handing them the keys to the kingdom.

The model can connect, scan, plan, and deploy — but every step that

changes state on a production-like server passes through a consent

gate the AI cannot self-approve. Discovery is read-only by design.

┌─────────────────┐          MCP / stdio          ┌────────────────────┐
│  AI client      │  ───────────────────────────► │  devops-mcp        │
│  (Claude /      │                               │                    │
│   Cursor / …)   │  ◄─────────────────────────── │  ssh2 / docker /   │
└─────────────────┘                               │  child_process     │
                                                  └────────┬───────────┘
                                                           │ SSH
                                                           ▼
                                                   ┌────────────────┐
                                                   │  Your VPS      │
                                                   └────────────────┘

⚡ First-time setup (read this once, do it once)

There are exactly four steps. Don't skip step 2.

1. Install

git clone .git devops-mcp
cd devops-mcp
npm install
npm run build

Requires Node ≥ 18.

2. Generate your elevation token and save it somewhere you won't lose it

# Linux / macOS
openssl rand -hex 24

# Windows PowerShell
$bytes = New-Object byte[] 24; (New-Object System.Security.Cryptography.RNGCryptoServiceProvider).GetBytes($bytes); [BitConverter]::ToString($bytes).Replace("-","").ToLower()

# Or, via Node
node -e "console.log(require('crypto').randomBytes(24).toString('hex'))"

You'll get something like `6ba329add30b19a5a347178f7e3705fdea0ac1aa66cb9274`.

> 🔑 SAVE THIS TOKEN BEFORE STEP 3.

>

> This token is the only thing standing between the AI and uncontrolled

> production access. The model never sees it. Whenever the AI wants to

> elevate to PROVISION/FULL mode, approve a destructive action, change a

> server's role, or write on a production-like server, you paste it once.

>

> Put it in a password manager. If you lose it:

> - You can hand-edit your MCP client's config to set a new one, or

> - You can ask the AI to call `rotate_consent_token` if you still have

> the old one (which is circular if you've lost both).

>

> There is no recovery flow. This is the gate; we don't ship a back door.

3. Add devops-mcp to your MCP client's config

For Claude Desktop, edit `claude_desktop_config.json`:

• Windows: `%APPDATA%\Claude\claude_desktop_config.json`
• macOS: `~/Library/Application Support/Claude/claude_desktop_config.json`
• Linux: `~/.config/Claude/claude_desktop_config.json`

Add (or merge into existing `mcpServers`):

{
  "mcpServers": {
    "devops-mcp": {
      "command": "node",
      "args": ["/absolute/path/to/devops-mcp/dist/index.js"],
      "env": {
        "DEVOPS_MCP_ELEVATION_TOKEN": "",
        "LOG_LEVEL": "info"
      }
    }
  }
}

The same structure works for Cursor, Windsurf, and any other MCP client —

the env block is the standard MCP way of passing secrets.

4. Fully quit and reopen your MCP client

Not "close the window." On Windows, that means system tray → Quit. The

elevation token is read at startup; the client has to restart for it to

take effect.

You're done. Next time you talk to the AI, say "add my server at …" and it

will walk you through.

Why this exists

Generic "run-any-command" MCP servers are dangerous on production boxes. A

model with full shell on a live server can — and will — restart the wrong

service, deploy onto an in-use port, `docker prune` a database volume, or

escalate itself to root because nothing told it not to.

devops-mcp draws a hard line between *reading* and *changing*:

• Reading is always allowed (within a read-only SAFE allowlist).
• Changing on a production-like server requires the human's token —

passed out-of-band, invisible to the model.

• Deploying a new project goes through a port-conflict check and a

reviewable script, not 40 ad-hoc commands.

Features

Access control

• Three-tier mode: `SAFE` (default, read-only allowlist), `PROVISION`

(system installs, 1 h default expiry), `FULL` (root, 30 min default expiry).

• Out-of-band consent token — elevation and approvals require a string

only the user has. The model literally cannot read it.

• Production write-gate — on `role: production` servers (or any server

the scanner flags as `productionLikely`), any non-SAFE command requires

`consentToken` + `acknowledgeProductionWrite: true`. Catastrophic

commands — ones that are irrecoverable without a backup (`rm` of a

non-temp path, `rm -rf /…`, `dd of=/dev/…`, `mkfs`, SQL `DROP

TABLE/DATABASE`, `docker rmi`, `docker volume rm`, `docker rm -v`,

`docker system prune`) — additionally require `backupVerified: true`.

Ordinary writes and recoverable ops (editing a file, `rm /tmp/scratch`,

bare `docker rm` of a container that can be recreated from its image)

do not need `backupVerified`. Refusals echo the **exact resolved

command**.

• Per-server policy — `allowedModes`, `blockedCommands`, `allowedPaths`,

`requireApproval` live in `config//server.json` and are

enforced on every SSH command.

• Required `role` — `add_server` will not let the AI silently default

the role; it must ask the user, and the response includes a

`roleConsequences` block the AI reads back to you.

• Token rotation — `rotate_consent_token` generates a fresh token

(defaults to dry-run; `apply: true` atomically updates your MCP client

config).

• Credential rotation — `update_server_credentials` rotates the

password, swaps the SSH key (incl. encrypted keys via `keyPassphrase`),

or migrates host/user/port without re-adding the server. The role,

restrictions, and scan profile stay intact. Closes any active session to

that server first, validates the new creds with a test connection, and

is consent-gated on production.

• AWS / EC2 `.pem` ready — onboard with the `.pem` file + username +

IP. Reference it in place (`externalKeyPath`) or copy it into the config

bundle (`keyFilePath`); add `keyPassphrase` only if the key is encrypted.

• Multiple simultaneous connections, keyed by serverId — the MCP holds

one SSH connection *per server*, not one global slot. Claude Desktop runs

a single shared MCP process across all your conversations; with a global

connection, two chats working on two servers would clobber each other

("session 1 on server A, session 2 connects to B, now A's commands hit

B"). Keyed connections let both coexist. `run_command` takes a `serverId`:

optional when exactly one server is connected, **required when two or

more are** — an ambiguous call is refused rather than guessed. Every

`run_command` response echoes `target.serverId` and `activeConnections`.

• Anti-target-drift — `run_command`, `set_mode`, and `get_current_mode`

responses all carry the connected server identity, so a conversation can

never silently end up operating the wrong box. `disconnect_server` takes

an optional `serverId` (or `"all"`).

• Live-session aware onboarding — `add_server` surfaces the

currently-connected server(s) in its response and tells the AI not to

auto-switch to the newly-added one without asking.

Discovery & planning

• Server discovery scan — read-only probe of OS, hardware, listening

ports, installed stack (docker / nginx / apache / node / pm2), running

containers, parsed nginx sites, systemd services. Output persisted as a

`ServerProfile`.

• Profile diff on reconnect — `diff_server_profile` re-scans and reports

what changed since the saved snapshot.

• Port-conflict awareness — `check_port_conflict` returns the listening

process + a free-port suggestion before deployment.

• Plan, don't fire — `plan_deployment` returns an idempotent bash

script the user reviews. The MCP does not execute it.

Safety hardening

• All command args shell-quoted before they hit the remote shell. No

more `sh -c ""` payloads splitting at the wrong shell level.

• Validator inspects args — `run_command({command:"ls", args:["; rm -rf /"]})`

no longer slips through with SAFE-mode `ls` validation.

• Quote-aware chain splitter — chains of read-only commands stay SAFE.

Diagnostic pipelines like `du -sh /opt/* ; echo --- ; df -h /` don't

require elevation. Each fragment is validated independently; the chain's

required mode is the max of its parts.

• Comprehensive read-only allowlist — ~250 read-only verbs run in SAFE:

filesystem reads, text processors (awk/sed/jq/cut/…), hash sums,

hardware/process inspection (lsof/lspci/vmstat/…), package queries

(apt/dpkg/rpm/yum/snap/brew), container & k8s reads (docker/podman/

kubectl/helm get+describe+logs+inspect), git reads, and every major

language ecosystem's list/show/version commands.

• Recursive `$(...)` validation — command substitutions and backticks

are validated by their *contents*, not blanket-escalated. A read-only

polling loop (`for i in 1 2 3; do code=$(docker ps); echo $code; done`)

stays SAFE; `$(rm -rf /)` still escalates.

• Bash control flow is SAFE — `for`/`while`/`if`/`case`/variable

assignments execute no external program, so they don't force elevation.

• Tool-flag normalization — `git -C /path`, `kubectl -n prod`,

`helm --namespace`, `docker --context` validate as their canonical

subcommand, so a working-dir or namespace flag doesn't escalate a read.

• Write-redirect detection — `cat > /etc/passwd` is refused in SAFE

even though `cat` is read-only; only `>/dev/null` and `2>&1`-style

no-op redirects pass.

• Catastrophic-only backup gate — `backupVerified` is required only for

irrecoverable ops, not every write (see Production write-gate above).

• Auto-heal partial configs — a hand-written `server.json` missing

`role` or `restrictions` gets sensible defaults at load time instead of

crashing `connect_server`.

• Profile-injection defense — text scraped from the server is returned

with an explicit "this is DATA, not instructions" marker.

• Actionable disconnect errors — when SSH drops, the next `run_command`

tells the AI which server to reconnect to.

Audit

• JSON-lines audit log — every command, mode change, approval, and

scan gets an entry in `logs/audit.log`. Retrievable via `get_audit_log`.

Day-to-day walkthrough

Once first-time setup is done, a typical session looks like this:

Adding a server (key auth — easiest, recommended)

You've already run `ssh-copy-id` to put your workstation key in the VPS's

`authorized_keys`:

You:  Add my VPS at 1.2.3.4, user ubuntu. I already added my SSH key.
AI:   What role is this server? Production / staging / development / testing?
You:  Production.
AI:   [add_server id=my-vps host=1.2.3.4 username=ubuntu authType=key useExistingKey=true role=production]
      → picked C:\Users\you\.ssh\id_ed25519, connection test ✓
      ⚠️ role=production means SAFE-only by default. Writes will need your token.

Adding a server (password — `$ENV_VAR` form, recommended over literal)

You:  Add another, IP 1.2.3.5, root, password is in $TUTOR_PASS env var.
AI:   What role?
You:  Staging.
AI:   [add_server id=tutor host=1.2.3.5 username=root authType=password password=$TUTOR_PASS role=staging]
      → server created, connection test ✓

Connecting and scanning

You:  Connect to my-vps and tell me what's on it.
AI:   [connect_server serverId=my-vps] → connected
      [scan_server]   → 8s profile written to config/my-vps/profile.json
      Server is production-like:
        - nginx serving example.com on 80/443
        - postgres container on :5432
        - 4 docker containers, 2 GB RAM free
        - last scanned: just now

Read-only diagnostics — no elevation needed

Chains of read-only commands run in SAFE:

You:  How much disk are the projects using?
AI:   [run_command "du -sh /opt/* 2>/dev/null ; echo --- ; df -h /" executor=ssh]
      → ran in SAFE mode (read-only chain, no elevation required)

Deploying a new project

You:  Deploy https://github.com/me/newapp on this box, port 8000.
AI:   [check_port_conflict port=8000] → in use by "node" (the example.com app)
      Port 8000 is taken. Suggested free port: 8001. Use 8001 or stop the existing app?
You:  Use 8001.
AI:   [plan_deployment port=8001 runtime=node ...] → returns a 26-line bash script
      Here's the script. Please review.
You:  Looks good. Run it.
AI:   [run_command ...] → refused: production write-gate.
      To run this I need your elevation token and confirmation that a backup exists.
You:  Token is . Yes, snapshot taken this morning.
AI:   [run_command consentToken= acknowledgeProductionWrite=true backupVerified=true]
      → ✓ deployed

Rotating credentials (the VPS password changed, or you swapped your SSH key)

You:  I rotated my-vps's root password. New one is in $MY_VPS_PASS_NEW.
AI:   my-vps is role=production — for the rotation I need your elevation token.
You:  Token is .
AI:   [update_server_credentials serverId=my-vps authType=password
        password=$MY_VPS_PASS_NEW consentToken=]
      → closed active SSH session (was connected), new creds tested ✓
      Reconnect with connect_server when ready.
You:  Connect.
AI:   [connect_server serverId=my-vps] → ✓

The role, restrictions, blocked-commands list, and scan profile are

preserved. Only the auth fields change.

Updating a server's role later

You:  Actually my-vps is staging now, not production.
AI:   This is a production-touching change, please confirm with the token.
You:  Token is .
AI:   [update_server serverId=my-vps role=staging applyRoleDefaults=true consentToken=]
      → role changed; allowedModes now [SAFE, PROVISION].

Rotating the token (when the old one has leaked, e.g. into chat)

You:  Generate a new elevation token and update Claude Desktop's config.
AI:   For verification, paste the current token.
You:  
AI:   [rotate_consent_token consentToken= apply=true]
      → 🔑 NEW TOKEN:   ← SAVE THIS NOW, in a password manager.
      Claude Desktop must be fully restarted for the new token to take effect.
      Until then, the OLD token still works on this running session.

Access modes

Elevation requires:

• `acknowledgeRisk: true` (the AI sets this)
• `consentToken: ""` (only you have it)

Downgrade is always allowed and instant. Sessions auto-expire back to SAFE.

Tool reference (32 tools)

Server lifecycle

Discovery (all read-only)

Execution & deployment

Mode, consent, audit

Authentication options (`add_server`)

Five ways to authenticate, picked by `authType` + which key/password field

you set:

The handler validates exactly one key source per call. Combining e.g.

`useExistingKey` and `keyFilePath` is refused with a clear error.

Any key path (`keyFilePath` / `externalKeyPath`) accepts an optional

`keyPassphrase` (literal or `$ENV_VAR`) for encrypted private keys.

AWS EC2 (the `.pem` case)

You get a `.pem` file, a username (`ubuntu`, `ec2-user`, `admin`, …), and

a public IP/DNS. Two ways:

// Reference the .pem where it sits (recommended — nothing copied)
{
  "id": "my-ec2", "host": "ec2-1-2-3-4.compute.amazonaws.com",
  "username": "ec2-user", "authType": "key",
  "externalKeyPath": "C:\\Users\\you\\Downloads\\my-key.pem",
  "role": "production"
}

// Or copy the .pem into the server's config folder (portable bundle)
{
  "id": "my-ec2", "host": "1.2.3.4", "username": "ubuntu",
  "authType": "key",
  "keyFilePath": "C:\\Users\\you\\Downloads\\my-key.pem",
  "role": "staging"
}

Most AWS keys have no passphrase — omit `keyPassphrase`. If yours is

encrypted, add `"keyPassphrase": "$MY_PEM_PASS"` and set that env var.

> Modern sshd + password auth: ssh2 needs `tryKeyboard: true` for sshd

> setups that use PAM (Ubuntu 22.04+, Debian 12, Amazon Linux 2023, RHEL 9).

> devops-mcp sets this automatically — passwords work even when the server

> has `PasswordAuthentication no` and only allows `keyboard-interactive`.

Server configuration on disk

Each server lives in its own folder:

config/
├── my-vps/
│   ├── server.json     # config (host, user, auth, role, restrictions)
│   ├── key.pem         # optional SSH private key (only if you used keyFilePath / privateKey)
│   └── profile.json    # written by scan_server
└── _example/
    └── server.json     # template

`server.json` example:

{
  "name": "Production Web",
  "host": "1.2.3.4",
  "port": 22,
  "username": "ubuntu",
  "authType": "key",
  "keyFile": "production.pem",
  "role": "production",
  "restrictions": {
    "allowedModes": ["SAFE"],
    "blockedCommands": ["rm -rf", "shutdown", "reboot", "dd"],
    "requireApproval": true
  },
  "description": "Main production web server"
}

For an `externalKeyPath` workflow (key stays in `~/.ssh/`):

{
  "name": "My VPS",
  "host": "1.2.3.4",
  "port": 22,
  "username": "ubuntu",
  "authType": "key",
  "externalKeyPath": "C:\\Users\\you\\.ssh\\id_ed25519",
  "role": "production"
}

For password auth (always prefer `$ENV_VAR`):

{
  "authType": "password",
  "password": "$MY_VPS_PASS"
}

`$NAME` is resolved to `process.env.NAME` at connection time. **Don't

commit literal passwords.**

If a `server.json` is missing `role` or has `restrictions: {}`, the MCP

fills in `role: "development"` defaults at load time (and warns in the

logs). This prevents `connect_server` from crashing on hand-written configs.

Environment variables

Security model

What devops-mcp protects against

• Model running blind on production — write commands on a server with

`role: production` or `productionLikely: true` are refused without the

consent token + explicit ack + (for catastrophic, irrecoverable ops)

`backupVerified`.

• Self-granted approvals — the model can't fabricate `consentToken`

because it never sees `DEVOPS_MCP_ELEVATION_TOKEN`.

• Argument injection — every arg to `run_command` is shell-quoted

before reaching the remote shell. Multi-line scripts inside `sh -c`

payloads survive intact.

• Smuggled commands in args — the validator inspects `command + args`

together, so `run_command({command:"ls", args:["; rm -rf /"]})` correctly

escalates to FULL.

• Over-broad chain refusals — chains of read-only commands stay SAFE.

Each fragment is independently validated; only the worst one wins.

• Prompt injection from scanned content — banners, container labels,

log lines are returned with an "untrusted data" marker. The tool response

tells the model: display, don't execute.

• Silent port collisions — `plan_deployment` and `check_port_conflict`

surface conflicts before deployment.

• Shell injection in deploy/configure helpers — every interpolated

value is shell-quoted; nginx configs are written via heredoc; branch

names and env-var keys are validated.

• Production write-gate refusals echo the exact command — so you can

read what was about to run, not the AI's paraphrase.

What devops-mcp does not do

• It does not sandbox the *connected* server. Once you're in FULL mode

with the token, the model can do anything the SSH user can.

• It does not encrypt the consent token at rest in your MCP client config.
• It does not back up your data — `backupVerified` is a human attestation,

not a check.

See SECURITY.md for the full threat model.

Token management

The elevation token is a static string stored in `DEVOPS_MCP_ELEVATION_TOKEN`

in your MCP client config. It does not expire.

What expires:

• `FULL` mode session — 30 min default
• `PROVISION` mode session — 1 h default
• Session SSH keys from `generate_ssh_key` — 30 min default

When a mode session times out it drops back to SAFE; the AI re-asks for

the same token to re-elevate.

Rotating the token

You:  Rotate the elevation token and update Claude Desktop's config.
AI:   For verification, paste the current token.
You:  
AI:   [rotate_consent_token consentToken= apply=true]
      → new token: 
      → 🔑 SAVE THIS NOW. Without it you're locked out of every write operation.
      → Fully quit and reopen Claude Desktop to activate it.

The MCP writes the new token atomically into your client config (only the

`DEVOPS_MCP_ELEVATION_TOKEN` key — everything else in the file is

preserved). The running MCP process keeps using the old token until

you restart the client.

If you lose both the old and new tokens between rotation and restart,

hand-edit the client config to set a new one — that's the recovery flow.

Project layout

src/
├── index.ts                       # MCP entry point (stdio)
├── types/                         # TypeScript types
├── core/
│   ├── logger.ts                  # JSON-lines structured logger + audit logger
│   ├── mode-manager.ts            # SAFE / PROVISION / FULL state machine
│   ├── command-validator.ts       # Allowlist + quote-aware chain splitter + wrapper-token scan
│   ├── server-config-manager.ts   # config//server.json + profile.json + auto-heal
│   ├── server-scanner.ts          # SAFE-mode discovery (read-only by design)
│   ├── ssh-key-manager.ts         # Session SSH keys with auto-expiry
│   └── approval-manager.ts        # Approval queue
├── executors/                     # Local / SSH / Docker — all shell-quote args
├── playbooks/                     # Provisioning playbooks (Docker, Nginx, …)
└── tools/
    ├── tool-schemas.ts            # Zod schemas + MCP tool definitions
    └── tool-handlers.ts           # The actual handlers

Development

npm run dev        # watch mode (tsx)
npm run build      # tsc → dist/
npm test           # vitest
npm run test:run   # vitest run (CI mode)
npm run lint       # eslint src/**/*.ts

Contributing

PRs welcome. See CONTRIBUTING.md.

When adding a new tool that *writes* on the connected server, make sure it

runs through `BaseExecutor.execute()` so the mode validator and the

production write-gate apply. Do not shell out directly from a handler,

and if you must interpolate a value into a shell command, use the

`shellQuote` helper in the executor — the historical bugs in this codebase

have all been quoting bugs.

License

MIT — see LICENSE.