MCP server

Let an AI assistant read and manage your containers — safely, on your terms.

logdeck mcp runs a Model Context Protocol server over stdio, so an assistant like Claude Desktop, Cursor, or Claude Code can query your containers, logs, events, and stats directly — and, when you opt in, restart a container or run a command. It is a thin layer over the same HTTP API the web UI and CLI use, so it adds no new server and no new way in: your existing API token decides what it can do.

Setup

You need the logdeck CLI installed (see the CLI reference) and a running LogDeck server. Add one entry to your MCP client's configuration:

{
  "mcpServers": {
    "logdeck": {
      "command": "logdeck",
      "args": ["mcp"],
      "env": {
        "LOGDECK_URL": "https://logdeck.example.com",
        "LOGDECK_TOKEN": "ldk_your_read_token"
      }
    }
  }
}

The server reads its connection from LOGDECK_URL and LOGDECK_TOKEN (or a saved CLI context). On startup it prints the server it is serving to stderr, and warns if the token is not a scoped ldk_ API token.

Capability follows the token

The MCP server never widens what your token can do. Mutations are enforced by the LogDeck server, not the MCP layer, so the safe path is the default:

A read-scoped token (ldk_, created in Settings → API access) can read logs, stats, events, and container details, but every action tool returns a permission error. This is what you hand to an assistant by default — safe by construction. See scoped API tokens.
An admin token can do everything an admin can do, including exec, environment variables, and settings — the same reach it has in the CLI and the web UI. You opt into that simply by choosing which token to configure.

Read tools

These are always available and are read-only.

list_containers

Containers across every host, including removed ones and their health state.

get_logs

Recent parsed logs for one container, with tail, level, regex, and time-range filters. Never follows.

search_logs

Regex search across many running containers, merged by timestamp.

inspect_container

The full inspect document for one container.

list_events

Docker/Podman events collected over a short bounded window.

container_stats / host_stats

Live CPU and memory per container, and per-host system stats.

list_images / list_volumes / list_networks

Images, volumes, and networks across hosts.

history_search / history_status / history_containers

Query the persisted log store: fast, indexed, cursor-paginated, and readable even for containers that no longer exist.

Action tools

These need an admin token. There are no flags to set: hand the assistant the token you want it to have, exactly as you would for the CLI.

start_container / stop_container / restart_container

Reversible lifecycle actions.

remove_container

Remove a container. Irreversible, and marked destructive so clients prompt harder.

run_command

Run one non-interactive command in a container and return separate stdout, stderr, and the exit code.

get_env / set_env

Read and replace a container's environment variables. Values commonly hold secrets, and a write recreates the container, so it restarts with a new ID.

get_settings / set_read_only / set_log_storage

Read settings, toggle server-wide read-only mode, and change log persistence and its retention caps.

set_docker_hosts / set_coolify_hosts

Replace the configured hosts. Each takes the complete list rather than merging, so read get_settings first.

set_auth / list_api_tokens / create_api_token / delete_api_token

Change authentication and manage API tokens. Disabling auth leaves the server open to anyone who can reach it.

Client confirmation prompts are a convenience, not a security boundary — the token scope is. A read-scoped token cannot use any action tool: the server rejects every mutation, and denies the env and settings surfaces outright. Destructive tools carry the protocol's destructive hint, so a well-behaved client prompts harder before running them.

Notes

Log tools default to a small tail and cap the number of lines returned, to stay within an assistant's context. history_search is cursor-paginated for walking further back.
run_command is non-interactive: it runs one command, returns its output and exit code, and does not attach a terminal. For an interactive shell, use the web terminal.