Container Host Aiops

Disclaimer: Community-maintained open-source project. Not affiliated with, endorsed by, or sponsored by Docker, Inc., Portainer.io, or any container-platform vendor. "Docker", "Portainer" and all product/trademark names belong to their respective owners. MIT licensed.

Governed AI-ops for non-orchestrator container hosts — the Docker Engine API (over a local unix socket or a TCP host), Portainer (its management API, which also proxies Docker), and Podman (a rootful/rootless service socket speaking the Docker-compatible API plus libpod-native endpoints) — with a built-in governance harness: a unified audit log over both MCP and CLI, a runaway/budget safety guard, and undo-token recording. It records every operation; it does not decide whether a write is permitted — that is the agent's or the account's call. Multi-platform by construction: a registry keyed by platform means a per-target platform field (docker / portainer / podman) selects the API shape, and another host family could be added later without touching the ops/CLI/MCP layers. Exercised against a live Docker Engine 27.5.1 daemon (doctor, overview, the three flagship analyses, and a governed stop_container with audit + undo recorded); the Portainer and Podman API paths are covered by the mock suite only.

What it does

Three flagship signature analyses, plus the guarded reads and writes around them:

• Restart-loop RCA — inspect containers for restart count + exit code, flag the crash-looping ones (restartCount over threshold, or restarting/dead, or a non-zero exit), and map each to a likely cause + action from the exit code (137 OOM/SIGKILL, 143 SIGTERM, 139 segfault, 127 bad entrypoint, …), with a tail of logs. Every ranking carries its numbers, not a black-box verdict.
• Resource-pressure analysis — a one-shot CPU%/memory% sample per running container vs its configured limits, flagging each "near" (≥ 80% of a threshold) or "over", with a recommendation (raise a limit, set a missing memory limit, scale out).
• Image & volume bloat — dangling images + dangling volumes + build cache from system/df, totalled into prune candidates with reclaimable bytes.

What works

• CLI (container-host-aiops ...): init, overview, container, image, volume, network, system, stack, pod, analyze, manage, secret, doctor, mcp.
• MCP server (container-host-aiops mcp or container-host-aiops-mcp): 38 tools (29 read, 9 write), every one wrapped with the bundled @governed_tool harness.
• Connection layer: Docker over a unix socket (httpx.HTTPTransport(uds=...)) or a TCP host; Portainer over HTTPS with an X-API-Key token that also proxies the Docker API of a managed endpoint; Podman over its rootful/rootless service socket (autodetected: $XDG_RUNTIME_DIR/podman/podman.sock first, then /run/podman/podman.sock) speaking the Docker-compat layer (paths reused wholesale) plus libpod-native endpoints. A local Docker/Podman socket needs no secret — the socket's file permissions are the boundary.
• Encrypted credentials: the Portainer API token lives in an encrypted store ~/.container-host-aiops/secrets.enc (Fernet + scrypt) — never plaintext on disk. Unlock with a master password from CONTAINER_HOST_AIOPS_MASTER_PASSWORD (MCP/CI) or an interactive prompt (CLI).
• Reversibility: mutating writes fetch the real before-state first and record a faithful inverse (stop↔start; update_container restores prior CPU/memory limits). Irreversible ops (remove_container, prune_images, prune_volumes, recreate_stack) capture the before-state for audit but declare no undo.
• Safety: every state-changing CLI op supports --dry-run and requires double confirmation; every write MCP tool takes a dry_run preview — and prune previews list what would be removed + reclaimable bytes before doing it.

What this tool does, and does not, decide