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: unified audit log, policy engine, token/runaway budget guard, undo-token recording, and graduated-autonomy risk tiers. 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.

Security: read-only mode

This tool is meant to be handed to an AI agent, so its safety story is enforced by the server rather than requested in a prompt:

export CONTAINER_HOST_READ_ONLY=1

With that set, the 9 write tools are never registered. An MCP client lists 29 tools instead of 38 — the writes are not hidden, not gated behind a flag, and not merely refused when called. They are absent from the session. A model cannot invoke a tool it was never offered, and cannot be argued into one.

That distinction is the whole point. A tool that exists but refuses still invites retry loops and "I'll describe the call instead" behaviour from smaller models, and it leaves a reviewer trusting a promise. An absent tool is a fact you can check: connect, list the tools, and see that the writes are not there.

Enforcement is two layers deep, so the switch cannot be sidestepped by changing entry point:

Read operations are unaffected, and every call is still audited to ~/.container-host-aiops/audit.db.

The read/write split is derived from each tool's declared risk_level, and a test asserts that this never disagrees with the [READ]/[WRITE] tag in the tool's own documentation — so a write can't quietly present itself as a read.

Running a smaller / local model? See agent-guardrails.md — it lists the guardrails this tool now enforces for you (so you don't spend prompt budget restating them) and gives a ready-made system prompt for what's left.

Capability matrix (38 MCP tools)

The three analyses accept injected data for offline analysis, or pull live from a configured target. list_endpoints/list_stacks/stack_detail require a portainer target; list_compose_stacks (Compose project rollup by the com.docker.compose.project label, with per-stack health) works on docker or podman; list_pods requires a podman target (Docker/Portainer have no pod concept).

Platform support matrix

Quick start

uv tool install container-host-aiops          # or: pipx install container-host-aiops
container-host-aiops init                      # wizard: add a Docker/Podman socket or Portainer target
container-host-aiops doctor                    # verify config, secrets, connectivity
container-host-aiops overview                  # one-shot host health
container-host-aiops analyze restart-loop      # crash-looping containers + cause/action
container-host-aiops container list --running  # running containers

Run as an MCP server (stdio):

export CONTAINER_HOST_AIOPS_MASTER_PASSWORD=...   # only needed for Portainer targets
container-host-aiops-mcp

Governance

Every MCP tool passes through the bundled @governed_tool harness:

• Audit — every call (params, result, status, duration, risk tier, approver, rationale) is logged to ~/.container-host-aiops/audit.db (relocatable via CONTAINER_HOST_AIOPS_HOME).
• Budget / runaway guard — token and call budgets trip a circuit breaker.
• Risk tiers — graduated autonomy; high-risk ops (remove / prune / recreate) can require a named approver.
• Undo recording — reversible writes record an inverse descriptor built from the fetched before-state.

Scope

This is the container-host member of the AIops-tools family (governed AI-ops with audit + budget + undo + risk tiers), for single-host Docker / Portainer / Podman. It is deliberately NOT for a cluster orchestrator, a hypervisor, a storage appliance, a backup product, or OT / industrial edge — those are separate tools/lines.

Missing a capability?

Coverage is intentionally a curated subset of the Docker Engine + Portainer + Podman (libpod) APIs. Missing a call, or want another container host family? Open an issue or PR — contributions welcome.

Verification status

• Docker — exercised against a live Docker Engine 27.5.1 daemon: doctor and overview connected over the local socket, the restart-loop and image/volume bloat analyses were run against real crash-looping containers and real reclaimable data, and a governed stop_container wrote a row to the audit DB and recorded a working undo descriptor.
• Portainer and Podman — mock-validated only. Those API paths are modelled from each project's public API shape and have not been exercised against a live server.
• The full checklist — what the mock suite guarantees, what the Docker run already satisfied, and what a Portainer/Podman run still has to prove — is in docs/VERIFICATION.md. container-host-aiops doctor is the fastest live check on any platform.