Chat with your Docker host. A Model Context Protocol server that turns any MCP‑capable AI — Claude Desktop, Cursor, Continue, Zed — into a natural‑language DevOps copilot for Docker & Docker Compose.
“Restart the container.” · “Why did crash — show me the last 100 log lines.” · “Deploy the stack in and confirm it’s healthy.”
Chat with your Docker host. An MCP server + interactive panel + terminal UI to manage Docker & Compose, with Prometheus metrics. By SoyRage Agency.
MCP normally means one server process per client, on the client's machine — backwards when the daemon you are managing lives on a server. Set DOCKER_MCP_HTTP=true and one instance serves every machine on your network over HTTP. Bearer token, loopback by default. → Run it over the network
🩺
"Is anything wrong?" in one call
host_health checks failing healthchecks, crash loops, containers down despite a restart policy, and disk pressure. find_restart_loops returns the offenders with the tail of their logs. find_unused_resources breaks down what is safe to clean. → Diagnostics
💬
Workflows your client offers you
MCP prompts — debug this container, audit my host, review my Compose file, free up disk. You no longer have to know which tool to ask for. → Guided workflows
Or open the live web panel straight away: npx -y -p @soyrageagency/docker-mcp docker-mcp-panel
Option A — standalone binary (no Node, no npm) · recommended for the CLI/TUI/panel
Since v1.1 you can install a single self-contained ragedocker executable — a Node runtime and the whole app fused into one file. Nothing else to install.
curl -fsSL https://raw.githubusercontent.com/soyrageagency/docker-mcp-server/main/scripts/install.sh | sh
Then just run ragedocker for an interactive menu, or ragedocker tui / ragedocker panel / ragedocker ia login. Re-run the installer any time to update. (How it's built and how to reuse this across repos: docs/DISTRIBUTION.md.)
Option B — via Node · configures Claude Desktop for you
Prefer the MCP-server-for-Claude-Desktop path? The installer clones the project, builds it, and configures Claude Desktop for you — no manual JSON editing. You need Git and Node.js ≥ 18.
npm run setup # builds + configures Claude Desktop
Then restart Claude Desktop and ask: “What Docker containers are running?” 🎉
The installer backs up your existing config and merges the entry, so any other MCP servers you have are preserved. Prefer to see the snippet without writing anything? node scripts/install.mjs --print.
🧭 What is this?
The Model Context Protocol (MCP) is an open standard that lets AI assistants talk to external tools through a well‑defined JSON‑RPC interface. Docker MCP Server is an MCP server that speaks that protocol over stdio and exposes your Docker host as a set of safe, richly‑described tools.
Point any MCP‑capable assistant at it and you can operate containers and Compose stacks in plain language — the model reads each tool’s schema, decides which to call, and reports the results back to you. No more memorising flags or copy‑pasting container IDs.
In one line: it’s the bridge between “I wish I could just tell my server what to do” and your actual Docker daemon.
💡 Why it exists
Day‑to‑day container work is a stream of small, repetitive commands:
docker ps -a
docker logs --tail 100 -f my-api
docker compose -f ./prod/compose.yaml up -d --build
docker stats my-api
Every one of those is trivial once you remember the exact syntax. The friction is the memorisation and the context‑switching. This server removes that friction by letting the AI do the translation, while keeping you in control through:
Read‑only mode for safe demos and production insight.
A container allowlist so the assistant can only touch what you allow.
Opt‑in exec so arbitrary in‑container commands are never available by accident.
Confirmation‑friendly design — destructive tools are clearly described so the model asks before it acts.
Built by SoyRage Agency for the self‑hosting and home‑lab community — and equally at home on a CI runner or a production VM behind read‑only mode.
🚀 Feature overview
Area
Capabilities
🔎 Insight
List containers · inspect full config · live CPU/memory/network stats · tail logs with time windows · list images / networks / volumes · host summary · disk usage.
Validate config · list services & health · deploy (up -d, optional --build) · tear down · restart · pull — via the official docker compose CLI.
🛡️ Safety
Global read‑only mode · container allowlist · opt‑in exec.
🔌 Transport
Local Unix socket · Windows named pipe · secured remote TCP + TLS.
🎨 Identity
ASCII welcome banner · about tool · MCP instructions that credit SoyRage Agency to the AI on connect.
🖥️ Interactive panel
Tabbed web dashboard: live monitoring, terminal with smart command suggestions, file explorer, snapshots & scheduled backups, networks/volumes, auto‑restart watchdog, alerts & log watch, inspect, search and lifecycle actions — with a demo mode.
⌨️ Terminal UI
A creative, lazydocker‑style TUI with a SoyRage welcome, live gauges and one‑key actions — zero curses dependencies.
📈 Monitoring
Built‑in Prometheus /metrics endpoint — scrape it from Prometheus, Grafana, Zabbix, VictoriaMetrics, …
🧩 Modular
Every capability is a toggleable plugin; enable exactly the surface you want via config.
🧱 Engineering
100% TypeScript, strict mode · one module per concern · tiny dependency surface · stderr‑only logging.
🖥️ The interactive panel
Beyond the conversational interface, the project ships a minimalist web dashboard (docker-mcp-panel) for when you want a fast, visual, point‑and‑click view of your host. It reuses the exact same configuration, Docker client and safety rails as the MCP server — so read‑only mode and the allowlist apply here too.
npm run build
ragedocker panel # → http://127.0.0.1:4600 (or: npm run panel)
npm run panel:demo # same, but with realistic mock data (no daemon needed)
One command for everything. Since v1.1 a single ragedocker launcher fronts the whole toolkit — run ragedocker with no arguments for an interactive menu, or:
ragedocker tui # the lazydocker-style terminal dashboard
ragedocker panel # the web panel + monitoring API
ragedocker mcp # the MCP server (Claude Desktop, Cursor, Continue…)
ragedocker ia login # sign in to Claude or ChatGPT (see below)
ragedocker doctor # check Docker, the AI and your configuration
The original docker-mcp-tui / docker-mcp-panel / docker-mcp-server binaries still work as aliases.
🔑 AI copilot — sign in once with ragedocker ia login
The TUI copilot (press a) and the panel's AI terminal need a model. A short wizard sets one up — pick Claude (Anthropic, used natively) or ChatGPT (OpenAI), paste your key, and it's verified before saving:
ragedocker ia login # set up your main AI account
ragedocker ia secundaria # add a second account to switch to
ragedocker ia use secondary
ragedocker ia list # show what's configured (keys masked)
ragedocker ia test # call the model once and report back
Keys are stored at ~/.ragedocker/ai.json, owner‑readable only (0600), and are never echoed or logged. The legacy DOCKER_MCP_AI_* environment variables still take precedence for unattended deployments.
Dashboard — live host stats, containers & images
One‑click log tailing
Terminal — run docker commands with smart suggestions (Tab / ↑↓ / Enter)
Snapshots & scheduled backups — with email/cloud delivery
File explorer + editor — view & edit files inside a container
Browse a container's filesystem
System — networks & volumes
Alerts & log watch — with a live auto‑restart watchdog
Tabbed layout — Overview, Terminal, Files, Backups, System and Alerts, with a live alert badge in the header.
Live monitoring + Trends — CPU‑load and memory‑used cards with meters, per‑container CPU % and memory bars, and historical sparkline charts (sampled every 5s) for CPU and memory over time.
⌨️ AI‑powered terminal — type docker … for typo‑tolerant, context‑aware completions (Tab/↑↓/Enter), or type a request in plain English (“why did web crash”) and pick Ask AI — the copilot proposes a command you review, then run. Commands are parsed to an argv array and spawned without a shell; a deny‑list blocks dangerous verbs and write verbs respect read‑only mode. AI uses any OpenAI‑compatible endpoint (OpenAI, Ollama, LM Studio…); without one it falls back to rule‑based suggestions.
📁 File explorer + editor — browse any container's filesystem (breadcrumbs, up‑navigation), open and edit text/.sh/config files, and Save back into the container. An AI edit button rewrites the file from a plain‑English instruction. Safe, shell‑free exec; edits are blocked in read‑only mode.
📸 Snapshots & scheduled backups — snapshot a container as an image (commit) or a filesystem export (.tar) to a chosen directory; schedule a daily backup (time, containers, type); a webhook forwards each backup to email, Google Drive or S3 via Zapier / Make / n8n.
🧩 System tab — networks and volumes at a glance; per‑container inspect details (env redacted, mounts, ports, restart policy) in the drawer.
♻️ Auto‑restart watchdog — flip the Auto toggle and a background watchdog restarts a container whenever it exits (respects read‑only mode).
🚨 Alerts & log watch — down/unhealthy containers, high CPU/memory, watchdog events, and error/warn lines scanned from recent logs.
Container grid — colour‑coded state dots, ports as chips, filter/search, auto‑refresh toggle, and per‑row actions (details · files · snapshot · start/stop/restart).
Prometheus /metrics — footer link exposes the scrape endpoint (see Monitoring).
Demo mode — DOCKER_MCP_PANEL_DEMO=true serves fabricated‑but‑realistic data (with gentle live jitter), perfect for previews and client demos with no daemon.
Zero UI dependencies — hand‑written HTML/CSS/JS served by a Node‑core HTTP server.
AI copilot (optional): set DOCKER_MCP_AI_ENDPOINT (+ DOCKER_MCP_AI_KEY, DOCKER_MCP_AI_MODEL) to power the AI terminal and AI file‑editing. Works with OpenAI, Ollama (http://localhost:11434/v1, free & local), LM Studio, or any OpenAI‑compatible API. Demo mode simulates it so you can try the UX offline.
🖼️ Regenerate the screenshots yourself with npm run shots (requires npx playwright install chromium).
🔒 Panel security & networking (VPN, IPs, ports)
The panel and terminal can control your host, so treat access like SSH.
Bind locally by default — the panel listens on 127.0.0.1:4600. Reach a remote host by tunnelling over a VPN — WireGuard or Tailscale — and browsing to the host's VPN IP. Do not port‑forward the panel to the public Internet.
If you must bind to the LAN, set DOCKER_MCP_PANEL_HOST=0.0.0.0. On startup the server prints every IP address it is reachable on and a warning, e.g.:
Panel ready at http://0.0.0.0:4600
reachable at http://127.0.0.1:4600
reachable at http://10.8.0.3:4600 ← your WireGuard IP
reachable at http://192.168.1.42:4600 ← your LAN IP
Panel is bound to ALL interfaces … keep it behind a VPN or an authenticated reverse proxy.
Change the port with DOCKER_MCP_PANEL_PORT (e.g. 8443) to avoid clashes or sit behind a reverse proxy.
Port‑forwarding / reverse proxy — if you expose it, put an authenticated proxy (Caddy/nginx/Traefik with Basic‑Auth or SSO + TLS) in front; never forward the raw port. Combine with DOCKER_MCP_READONLY=true for view‑only deployments, and DOCKER_MCP_PANEL_TERMINAL=false to disable the command runner.
⌨️ The terminal UI (TUI)
Prefer the terminal? Launch docker-mcp-tui — a creative, lazydocker‑style dashboard that opens with a SoyRage Agency welcome and then drops you into a live, keyboard‑driven view. It’s hand‑rolled ANSI (no curses library), so it adds zero dependencies.
ragedocker tui # → interactive terminal dashboard (or: npm run tui)
npm run tui:demo # same, with realistic mock data (no daemon needed)
A warm welcome — “thank you for using our repository ⭐”
Live dashboard — gauges, details & one‑key actions
AI copilot — give an order in plain English
AI copilot. Press a (or :) to open the copilot bar and type what you want — “restart the web container”, “stop the backup job”, “show every container”. The copilot turns it into the exact docker command, shows it to you, and runs it only after you confirm with y. It speaks to any OpenAI‑compatible endpoint (DOCKER_MCP_AI_ENDPOINT / DOCKER_MCP_AI_KEY / DOCKER_MCP_AI_MODEL) and falls back to a built‑in rule engine in demo mode, so you can try it with zero setup.
Action menu. Press Enter (or m) over any container to open a menu of everything you can do to it — logs, inspect, shell, start/stop/restart, snapshot, restore, attach a volume, edit compose, copy id — arrow‑key navigable, so you never have to memorise a key. Every action also has a one‑key shortcut for power users.
Animated status. Containers no longer show a plain red/green dot: healthy ones breathe (●↔◉), starting/restarting ones spin (⠋⠙⠹…), unhealthy ones blink (▲), paused show ⏸, exited ○ — a glance across the list reads like a heartbeat monitor.
Keys:Enter/mmenu · ↑/↓ (or j/k) navigate · g/G top/bottom · / filter · o sort (O reverse) · space pause live updates · l logs (f follow, +/- tail) · i inspect · e shell · c edit compose · y copy id · n/v/A networks/volumes/alerts · a AI · S/s/R lifecycle · b snapshot · B restore · V attach volume · u what's new · ? help · q quit.
Live CPU/memory gauges refresh automatically; the details pane shows lazydocker‑level info (health, uptime, command, restart policy, networks, mounts, env), lifecycle actions play a spinner animation, and read‑only mode hides the mutating keys (the copilot still explains, but never runs a mutating command).
Update notices. When a new version is published, both the TUI (a banner + u for the changelog) and the web panel (a dismissable banner) tell you — read from a single updates.json in the repo. It's silent on any failure and can be turned off with DOCKER_MCP_NO_UPDATE_CHECK=1.
📈 Monitoring: Prometheus, Zabbix & more
The panel doubles as a metrics exporter. It serves a standard Prometheus text endpoint at /metrics, so your Docker host becomes a first‑class monitoring target with no extra agent.
npm run panel # metrics on by default
curl http://127.0.0.1:4600/metrics
Exposed series (labelled by name, state, image where relevant):
Use an HTTP agent item pointed at http://your-host:4600/metrics, then add
dependent items with the Prometheus pattern preprocessing step, e.g.
dockermcp_containers_running or
dockermcp_container_cpu_percent{name="api"}. Grafana, Grafana Agent,
VictoriaMetrics and Netdata can scrape the same endpoint.
Turn the exporter off with DOCKER_MCP_PANEL_METRICS=false if you only want the UI.
🧩 Modular plugin architecture
The server is assembled from independent plugins, each owning one capability group. Which plugins load is driven entirely by configuration, so you can expose exactly the surface you want — from insight only to the full toolbox — without touching code. This also makes the project easy to extend.
🔒 The about plugin is locked — it carries the SoyRage Agency identity and cannot be disabled.
Toggle plugins via environment variables or the config file:
# Expose ONLY read-only insight (a safe, curated surface)
DOCKER_MCP_PLUGINS=containers,logs,images,system
# Load everything except container lifecycle
DOCKER_MCP_DISABLED_PLUGINS=lifecycle
Ask the assistant “list the plugins” any time to see what’s enabled.
Config file
For a reproducible setup, drop a docker-mcp.config.json in the project root (or point DOCKER_MCP_CONFIG at one). Environment variables always override it. See examples/docker-mcp.config.json:
┌──────────────────────────────────────────────┐
You ◀──────▶ │ AI assistant (Claude / Cursor / Continue …) │
└───────────────────────┬──────────────────────┘
stdio · JSON‑RPC (MCP)
┌───────────────────────▼──────────────────────┐
│ Docker MCP Server │
│ │
│ 1. Client sends `initialize` → server │
│ replies with tool schemas + SoyRage │
│ `instructions` (identity & welcome). │
│ 2. Model picks a tool and sends `tools/call`.│
│ 3. Server executes it against Docker and │
│ returns human‑readable text. │
└───────────┬───────────────────────┬──────────┘
Engine API │ spawn │ docker compose
┌───────────▼───────────┐ ┌──────────▼──────────┐
│ Docker Engine │ │ Compose plugin │
└───────────────────────┘ └─────────────────────┘
Engine operations (containers, images, stats, logs, system info) use the Docker Engine API through dockerode.
Compose operations shell out to the official docker compose CLI with a shell‑free, fully argument‑quoted spawn (no string interpolation, no injection surface).
stdout is sacred: it carries only the JSON‑RPC stream. Every log line goes to stderr.
✅ Requirements
Requirement
Notes
Node.js ≥ 18
ES modules + modern APIs. Node 20+ recommended.
A reachable Docker Engine
Local socket by default; remote TCP/TLS supported.
docker CLI on PATH
Only needed for the Compose tools. Insight/lifecycle tools work without it.
An MCP client
Claude Desktop, Cursor, Continue, Zed, or the MCP Inspector.
📦 Installation
# 1. Clone
git clone https://github.com/<your-user>/docker-mcp-server.git
cd docker-mcp-server
# 2. Install dependencies
npm install
# 3. Build to dist/
npm run build
Kick the tyres with the official MCP Inspector (no AI client required):
npm run inspect
This opens a UI where you can list tools and call them by hand — perfect for verifying your Docker connection before wiring up an assistant.
No install step needed: npx fetches the package on first run and keeps it up to date. Restart Claude Desktop and ask: “What containers are running?”
Cursor / Continue / Zed
Any MCP‑capable client works the same way — register a stdio server whose command is npx with the arguments -y @soyrageagency/docker-mcp, passing the same environment variables. Consult your client’s MCP documentation for the exact config location; the server block is identical.
🌐 Run it over the network (HTTP)
By default an MCP server talks over stdio: your AI client starts a copy of
it as a child process, on your machine. That is fine for a laptop tool and
awkward for a homelab, where the Docker daemon lives on a server — and where
you probably want your desktop and your laptop talking to the same thing.
Set one variable and it serves Streamable HTTP instead:
Read this before you expose it. The Docker socket is root on your host, so
anyone who can reach this port can do anything to it. It binds loopback by
default and warns loudly if you start it without a token. Put it behind your
VPN, and consider pairing it with DOCKER_MCP_READONLY=true or a container
allowlist for anything you do not fully trust.
There is also GET /health, which needs no token, for container healthchecks.
🩺 Diagnostics
Three read-only tools aimed at the questions you actually ask, rather than at
the API surface:
host_health — one call that answers is anything wrong? Failing
healthchecks, containers crash-looping, containers that are down despite a
restart policy (the failure nobody notices), and disk pressure from
reclaimable images and volumes. Findings ranked by severity, not a data dump.
find_restart_loops — the containers that keep dying, with the exit code
and the tail of their logs, which is where the reason almost always is. It
saves the usual ps → inspect → logs round trip that takes three messages.
find_unused_resources — what is safe to clean and what it would return,
split into images, volumes and long-dead containers. It refuses to be glib
about volumes: an unreferenced volume is very often the database of a stack you
stopped on purpose, and docker volume prune is not undoable.
All three are read-only, so they stay available with DOCKER_MCP_READONLY=true.
💬 Guided workflows (prompts & resources)
Tools only answer a question you already knew how to ask. MCP prompts are
the other half: your client lists them, so the workflow is discoverable without
knowing which tool to reach for or in what order.
Prompt
What it does
debug-container
Works out why a container is unhealthy or restarting, from its logs, config and environment — and says so when the logs do not actually support a conclusion.
audit-host
Read-only sweep of health, waste and risky configuration. Narrow it with health, disk or security.
review-compose
Reads your Compose file and points out what bites later: latest tags, missing healthchecks, state not in a volume, ports on 0.0.0.0, secrets in the file.
free-up-disk
Finds reclaimable space and ranks it by space returned ÷ risk.
Each prompt also tells the assistant what a good answer looks like: verdict
first, findings ranked, no padding, and an explicit instruction not to change
anything.
Resources expose the host as attachable context, so the model can be handed
the current picture instead of spending three tool calls rebuilding it:
docker://host/overview — containers, images and disk usage in one snapshot.
docker://server/capabilities — which plugins are loaded, whether the server
is read-only, whether exec is allowed, and what the allowlist permits.
⚙️ Configuration reference
Every setting is an environment variable. A local .env file (next to package.json) is loaded automatically at startup; values already present in the process environment always win, so your MCP client can override the file. See .env.example for a commented template.
Directory containing ca.pem, cert.pem, key.pem for a TLS‑secured remote daemon.
DOCKER_TLS_VERIFY
false
1/true to verify the daemon certificate (recommended for remote hosts).
DOCKER_MCP_READONLY
false
When true, all state‑changing tools are hidden — the server exposes insight only.
DOCKER_MCP_ALLOW_EXEC
false
When true, registers the exec_in_container tool (arbitrary in‑container commands).
DOCKER_MCP_CONTAINER_ALLOWLIST
—
Comma‑separated container names or prefixes the server may operate on. Empty = all. Prefix matching means web covers web-1, web-2.
DOCKER_MCP_DEFAULT_LOG_TAIL
200
Default number of log lines returned by container_logs when tail is omitted.
DOCKER_MCP_COMPOSE_CWD
process cwd
Base directory used to resolve relative Compose file paths.
DOCKER_MCP_LOG_LEVEL
info
Diagnostic verbosity written to stderr: debug | info | warn | error.
DOCKER_MCP_PLUGINS
—
Load only these plugins (comma‑separated). Empty = all.
DOCKER_MCP_DISABLED_PLUGINS
—
Disable these plugins (comma‑separated). about is locked.
DOCKER_MCP_PANEL_HOST
127.0.0.1
Bind address for the interactive panel.
DOCKER_MCP_PANEL_PORT
4600
Port for the interactive panel.
DOCKER_MCP_PANEL_DEMO
false
Serve fabricated demo data in the panel/TUI.
DOCKER_MCP_PANEL_METRICS
true
Expose the Prometheus /metrics endpoint.
DOCKER_MCP_PANEL_TERMINAL
true
Enable the in‑panel command terminal.
DOCKER_MCP_BACKUP_DIR
./snapshots
Directory for container snapshots/exports.
DOCKER_MCP_BACKUP_WEBHOOK
—
Webhook called after each backup (email/cloud bridge).
DOCKER_MCP_AI_ENDPOINT
—
OpenAI‑compatible base URL for the AI copilot (empty = off).
DOCKER_MCP_AI_KEY
—
Bearer key for the AI endpoint.
DOCKER_MCP_AI_MODEL
gpt-4o-mini
Model name for the AI endpoint.
DOCKER_MCP_CONFIG
docker-mcp.config.json
Path to the optional JSON config file.
Boolean parsing: any of 1, true, yes, on (case‑insensitive) counts as true. A JSON config file provides defaults for all of the above — see Config file.
🌐 Connecting to remote / TLS daemons
Manage a Docker host over the network by pointing DOCKER_HOST at its TCP endpoint. For anything beyond localhost, always use TLS.
When TLS is enabled the server reads ca.pem, cert.pem and key.pem from DOCKER_CERT_PATH and connects over HTTPS (default port 2376; plain TCP defaults to 2375).
🛡️ Security model
This server can control your infrastructure, so it ships with defence‑in‑depth defaults. You decide how much power to grant.
Control
What it does
Recommended for
Read‑only mode (DOCKER_MCP_READONLY=true)
Hides every state‑changing tool. The model literally cannot see stop, remove, deploy_stack, etc.
Restricts all container tools to matching names/prefixes. Anything else returns a clear “not allowed” error.
Multi‑tenant hosts, “manage the app, never the database”.
Opt‑in exec (DOCKER_MCP_ALLOW_EXEC)
The powerful exec_in_container tool is not registered unless you enable it.
Keep disabled unless you specifically need it.
Shell‑free Compose
Compose commands are spawned as argument arrays — no shell, no interpolation.
Always on.
Graceful errors
A failing tool returns an isError text result instead of crashing the transport, so a bad call never takes the session down.
Always on.
Safety recipes
# Give a live demo with zero risk of mutation
DOCKER_MCP_READONLY=true
# Let the AI manage only the app tier, never data stores
DOCKER_MCP_CONTAINER_ALLOWLIST=web,api,worker
# Never allow shelling into containers (this is the default)
DOCKER_MCP_ALLOW_EXEC=false
⚠️ Principle of least privilege. Start read‑only, add an allowlist, and only enable writes/exec once you trust the setup. Treat the assistant as a very fast junior engineer: helpful, but you sign off on the destructive stuff.
🧰 Complete tool reference
Tools marked W change state and are hidden when DOCKER_MCP_READONLY=true.
exec_in_container is additionally hidden unless DOCKER_MCP_ALLOW_EXEC=true.
Identity
Tool
Parameters
Description
about
—
Returns the welcome banner, version and credits.
list_plugins
—
Lists the modular capability plugins and whether each is enabled.
Insight (read‑only)
Tool
Parameters
Description
list_containers
all?: boolean
List containers with state, image, status and published ports. all includes stopped ones.
inspect_container
container: string
Full low‑level config for one container (env, mounts, network, restart policy, health) plus a readable summary.
container_stats
container: string
One‑shot snapshot of live CPU %, memory usage/limit and network RX/TX.
docker compose up -d --remove-orphans — deploy/refresh a stack, optionally rebuilding and scoped to services.
compose_down
file, project?, removeVolumes?: boolean
Stop and remove a stack. removeVolumes also deletes named volumes (destructive).
compose_restart
file, project?, services?: string[]
Restart all or selected services.
compose_pull
file, project?, services?: string[]
Pull the latest images for a stack (pair with deploy_stack for a rolling update).
💬 Example conversations
Natural‑language prompts and the tools the model will typically reach for:
You say…
The assistant calls…
“What’s running right now?”
list_containers
“Show me everything, including stopped ones.”
list_containers { all: true }
“Why did api crash? Last 100 lines.”
container_logs { container: "api", tail: 100 }
“Anything in the web logs from the last 15 minutes?”
container_logs { container: "web", since: "15m" }
“Is db using a lot of memory?”
container_stats { container: "db" }
“Restart nginx.”
restart_container { container: "nginx" }
“Deploy the stack in ./prod and rebuild.”
deploy_stack { file: "./prod", build: true }
“Which services are up in the demo stack?”
compose_ps { file: "examples/demo-stack" }
“How much disk is Docker using?”
disk_usage
“Who built this integration?”
about
Want a stack to practise on? examples/demo-stack/compose.yaml spins up nginx + redis. Try: “Deploy the demo stack, then show me its services and the web logs.”
stdout is reserved for the JSON‑RPC protocol stream; every diagnostic goes to stderr. Violating this corrupts the MCP connection — the logger enforces it.
No shell interpolation. Compose commands are spawned with an argument array, never a shell string, eliminating command‑injection risk.
Fail soft. A handler that throws returns a clean isError text result the model can read and recover from, instead of tearing down the transport.
One concern per module. Tools are grouped by capability; each group is a small, focused file that receives its dependencies explicitly (no globals).
Tiny dependency surface. A hand‑rolled .env loader keeps dotenv out; only @modelcontextprotocol/sdk, dockerode and zod are runtime dependencies.
Safety by construction. Read‑only mode and the allowlist are checked at the boundary, so an unsafe call can’t slip through a forgotten branch.
🧪 Development
npm run dev # hot‑reload the MCP server with tsx
npm run typecheck # strict type check, no emit
npm run build # compile to dist/ and copy panel assets
npm run start # run the built MCP server
npm run inspect # launch the MCP Inspector against the built server
npm run panel # run the interactive panel (with /metrics)
npm run panel:dev # hot‑reload the panel with tsx
npm run panel:demo # run the panel with demo data
npm run tui # run the terminal UI
npm run tui:demo # run the terminal UI with demo data
npm run shots # regenerate panel screenshots (needs Playwright chromium)
npm run clean # remove dist/
Coding standards: TypeScript strict with noUnusedLocals, noUnusedParameters, noImplicitReturns and noFallthroughCasesInSwitch.
Continuous integration:.github/workflows/ci.yml runs on every push/PR — install, type‑check, build, then the deep end‑to‑end suite (node scripts/deep-test.mjs, 70 checks) against a headless demo panel. Run it locally anytime:
npm run build
DOCKER_MCP_PANEL_DEMO=true DOCKER_MCP_PANEL_PORT=4600 node dist/panel/index.js &
node scripts/deep-test.mjs
🩺 Troubleshooting & FAQ
The server started but couldn’t connect to Docker. Check that:
Docker Desktop / the daemon is running.
DOCKER_HOST is correct for your platform (empty = default socket).
On Linux, your user can access the socket (docker group) or you’re running with sufficient permissions.
The server intentionally keeps running so tool calls return a friendly error inside your chat client instead of crashing.
The Compose tools shell out to docker compose. Install Docker Desktop or the docker-compose-plugin, and make sure docker is on the PATH of the environment your MCP client launches the server in.
You’re probably in read‑only mode. Set DOCKER_MCP_READONLY=false (the default) and restart your MCP client so it re‑reads the tool list.
DOCKER_MCP_CONTAINER_ALLOWLIST is set and the target doesn’t match. Add its name/prefix to the list, or clear the variable to allow all.
No. This server talks only to your Docker daemon and your MCP client over local stdio. It makes no outbound network calls of its own.
🗺️ Roadmap
The full roadmap — what is shipped, what is next, and what is deliberately
not planned — lives in ROADMAP.md.
The short version: a multi-arch ghcr.io image and a socket-proxy recipe, then
finer-grained permissions than the current read-only/read-write switch. Image
update checking and metrics history are further out because doing them badly
means confidently wrong answers.
Explicitly not planned: replacing Portainer, autonomous action without
confirmation, and telemetry of any kind.
🤝 Contributing
Contributions are welcome! Please:
Open an issue describing the change before large PRs.
Keep the stderr‑only logging and shell‑free Compose invariants intact.
Run npm run typecheck && npm run build before submitting.
🏢 About SoyRage Agency
SoyRage Agency is a full‑stack development & infrastructure studio based in Valencia, Spain, building tools where DevOps meets AI. We craft polished, production‑minded software for developers and the self‑hosting community.
If this project is useful to you, a ⭐ on the repo and a link back to soyrage.es genuinely help us keep building in the open. Thank you! 🙌
🧰 More from the SoyRage self‑hosting suite
Docker MCP Server is part of a family of open‑source infrastructure tools built with the same care — same design language, same safety‑first defaults, same "chat with your infra" philosophy:
Leaving vSphere after the Broadcom price hikes? Inventory vCenter, score compatibility, estimate cost & time, plan disk conversion and export a professional PDF assessment.
Living infrastructure documentation — agentless discovery that auto-generates a network diagram, inventory, VLAN & service-dependency maps, and tells you what changed since last time.
Ainda não há lista de ferramentas. Servidores remotos recebem uma após a primeira verificação de saúde.
Perguntas frequentes
Perguntas
Como instalo Docker?
Copie nesta página a configuração stdio para Cursor, Claude Desktop, Claude Code ou VS Code. Ela usa a instalação do publicador, não um proxy do PlugMCP.
Docker precisa de uma chave de API?
Este registro não lista nenhuma variável de ambiente obrigatória.
Qual transporte Docker usa?
Docker roda localmente via stdio.
Quando Docker foi verificado pela última vez?
Última verificação: 2026-10-28. Chat with your Docker host. An MCP server + interactive panel + terminal UI to manage Docker & Compose, with Prometheus metrics. By SoyRage Agency.