Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Self-Hosted AI Fleet Ops Bundle

    1

    A rigid evening runbook for managing self-hosted AI fleets with multi-box safety and verification ladders.

    Free

    Security scanned
    Self-Hosted AI Fleet Ops Bundle

    Self-Hosted AI Fleet Ops Bundle

    Example session with this skill installed

    I need to update the model on box-01 to Llama 3 but I have several agents running on cron. Help me run the preflight and backup.

    • Read your context and instructions
    • Compiled the self-hosted ai fleet

    I have initiated the Stage 1 Preflight and Stage 2 Backup sequence for box-01. I have audited the system resources and prepared a timestamped backup directory with a rollback card.

    Self-Hosted Ops: Preflight & Backup Report (box-01)

    CategoryStatusDetails
    RAM CheckPASS42GB Available. Llama 3 (8B) requires ~5.5GB. Headroom is sufficient.
    Disk SpacePASS140GB free on /var/lib/docker. Sufficient for weights + backup image.
    Co-tenantsALERTFound agent-triage-main (cron every 5m) and vllm-watchdog (timer).
    Port MapOKPort 8000 (vLLM) owned by PID 1204. No collisions on target 8001.
    Backup DirCREATED/opt/ai-fleet/backups/vllm.preupd-20231027-1930 (Mode 700)
    ArtifactsCAPTUREDdocker inspect, unit file, and ps auxww snapshot saved.

    Rollback Card Preview

    
    # Emergency Restore Sequence (Target: <5 mins)
    docker stop vllm-llama3 || true
    docker tag vllm-llama2:preupd-20231027-1930 vllm-production:latest
    docker run -d --name vllm-llama2 [args-from-snapshot.json]
    
    # Verify Rung 1
    curl http://localhost:8000/v1/models
    

    Next steps
    1.

    Wait out the cycle: The agent-triage-main job runs in 2 minutes. Wait for it to clear before stopping the container.
    2. Execute Change: Deploy the Llama 3 container once the pgrep for the agent returns empty.
    3. Climb the Ladder: Do not skip the LAN check (Rung 2) after the local health check passes.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Running one model on one box is a weekend project. Running a fleet — multiple machines, co-tenant production services, cron-driven agents, watchdogs, and a steady drip of model drops — is a discipline. The difference between the two is not hardware. It is operating procedure, written down and followed every single time, even on the nights you are tired and the change looks trivial.

    This bundle is the complete evening runbook: the preflight checklist, the backup conventions, the change discipline, the verification ladder, and the soak rules that keep a self-hosted AI stack stable for months. It is the meta-skill underneath every model swap and UI deployment — compiled from real fleet operations where agents run on schedules and a bad change costs trust, not just uptime. If you operate more than one box, or one box with more than one tenant, this is the runbook you wish you had written before the first outage.

    What you get

    • The seven-stage evening runbook — preflight, backup, wait-out, change, verify, breadcrumb, soak: the full arc of any production change, in order, with the exact commands for each stage.
    • Co-tenant production safety — how to map cron and watchdog cycles on a shared box, wait out the schedules, and keep your serving blip under one interval so scheduled agents never notice.
    • Backup conventions that survive operator churn — dated .preupd artifact naming, rollback cards written before the change and testable offline, and a five-minute recovery target.
    • The verification ladder — loopback, LAN, end-to-end: why every outage story starts with "we checked rung 1 and declared victory," and the per-account, per-lane E2E checks that listing endpoints cannot fake.
    • Breadcrumbs and the 48-hour soak — notice files for the next operator, one-variable-at-a-time change discipline, and the soak rule that keeps every regression attributable.

    One evening of discipline, learned from other people's 2 a.m. pages.

    Covers wiring and configs only. Requires separate procedures for data volumes, databases, and model artifacts. Quiet-polling for cron gaps remains inherently race-prone.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 1 day ago

    • Free with an Agensi account

    Permissions

    Terminal / Shell
    Network Access
    Write Files
    Listed1 day ago

    What's inside

    Frequently Asked Questions

    This skill provides a structured framework for high-stakes updates, preventing common disasters like 'silent' agent failures, RAM/swap thrashing during model swaps, and Docker-induced firewall bypasses.

    The skill is designed for Linux-based environments using Docker and systemd; it works with any LLM engine or agent framework that communicates via standard networking protocols.

    The 'Verification Ladder' is a three-tier testing sequence that validates your changes at the localhost, LAN, and end-to-end consumer levels before finalizing any deployment.

    You receive the complete operational runbook, diagnostic logic for preflight checks, and the structured 'Rollback Card' templates for offline recovery.

    While the skill manages container networking, service states, and configuration logic, it does not handle application-level database backups or active model training checkpoints.

    It is highly effective for multi-box setups, providing the necessary discipline to ensure that shared-resource hosts load large models sequentially and maintain network attribution.

    New in DevOps & Cloud