Repository navigation
Releases
Releases · falkenslab/docker-mcp
v0.4.0
Changed
- Stopping a lab doesn't ask any more.
lab_downstops a lab without asking the person, likelab_upstarts it: its files stay andlab_upbrings it back, so an agent can clean up the labs it started even when nobody is there to answer (#11). Resetting a lab (lab_reset) still asks.
Clients that gated lab_down on their side (such as the docker-lab extension's serverConfirmTools) can drop it from that list.
v0.3.0
What's new
- Confirmations in the person's language (#9). The form the server shows before a removal or a dangerous container follows the locale the app starts it with:
LC_ALL, thenLC_MESSAGES, thenLANG(es_ES.UTF-8is Spanish), as agent-kit passes it to its extensions' servers. English and Spanish; English without a locale or for another language. What the model reads stays English.server_inforeportslanguage.
Fixes
- A confirmation no longer expires after a minute (#10). It was sent with the MCP SDK's default 60-second request timeout, so a person who took longer got a refusal reported as "the person couldn't be asked". It now waits up to an hour, and a question that does expire says so.
What clients must change
Nothing. To get Spanish forms, start the server with a Spanish locale (agent-kit does it already).
Documentation: https://falkenslab.github.io/docker-mcp/ · Releases: https://falkenslab.github.io/docker-mcp/releases
v0.2.0
What's new
- You confirm, not the agent. Before removing a container, an image, a volume or a network,
prune(saying how much it would free),compose_downwith volumes,lab_downandlab_reset, and before a container with dangerous options, the server asks the person itself through MCP elicitation; the answer never reaches the model (ADR-007). - Isolated one-off containers.
run_containertakescpus,memoryandpidslimits,readOnly,noNewPrivilegesandlabels;network: "none"cuts a container off; a foreground run is stopped and removed whentimeoutSecondsruns out.list_containerstakes alabelfilter. - Podman, through its Docker-compatible socket (documented), and
npm run check:labsto check what labs rely on against any engine.
What clients must change
- A client without elicitation now gets a refusal from those tools. Claude Code and Codex support it. An app that asks before every call starts the server with
DOCKER_MCP_CONFIRM=client. run_containerno longer takesconfirmedDangerous(it's ignored if passed). Skills that told the model to set it should stop.server_inforeportsconfirm.
Fixes
- On Windows, a foreground run killed for its time didn't answer until the container ended.
- The docs named an extension that doesn't exist: it's
docker-lab, in falkenslab/agent-extensions.
Documentation: https://falkenslab.github.io/docker-mcp/ · Releases: https://falkenslab.github.io/docker-mcp/releases
v0.1.0
The first release of docker-mcp: Docker for AI agents, as an MCP server run with npx -y @falkenslab/docker-mcp.
- 47 tools over containers, images, volumes, networks, Compose, the system and labs, with structured results: among them what containers consume and run (
container_stats,container_processes), building a Compose project without starting it (compose_build), and pushing an image with the person's owndocker login(push_image), never with credentials passed to the agent. - Modes chosen at startup:
read(only reads),lab(only labs, under a safety policy, seeing only their containers) andfull(everything, with destructive tools marked and dangerous container options needing the person's yes); areas to narrow a mode. - Labs: a folder with a compose file and a check, refused when it breaks the policy, run as its own Compose project with limits and, unless allowed, without internet.
- Docker 24 or later as a requirement: checked and explained, never installed.
Documentation: https://falkenslab.github.io/docker-mcp/