Deployment packages: getting the local AI running
Six areas: Traefik, Harbor, cloudflared, dsh, k3d/k3s and Grafana. Scope agreed first, then a quote. Delivered as configuration, documentation and verification steps.
Updated 2026-11-16
What this is
A one-time project, not long-term operations and not a course.
- It covers: work in your machine and network that takes one area from "I know I should do this" to "it runs and I can maintain it", with configuration and documentation left behind.
- It does not cover: taking over accounts, staying on site, 24/7 on-call, unlimited revisions, or a guarantee for third-party services or model performance.
- Price: quoted by scope. Hardware, current setup, migration, timing and delivery depth all matter.
- Payment: no online payment yet. Once scope is agreed, a written quote covers the work, timing, payment and cancellation terms.
If you only want to know whether the machine is worth the effort, use the free assessment. No payment needed.
The six areas
Traefik reverse proxy
With several services, ports, hostnames and certificates scatter. Traefik pulls the entry point into one place, routes by hostname or path, and renews certificates on its own.
Delivered: a route table, configuration, certificate approach, debugging for the usual errors, and how to add an entrance without restarting everything. It does not handle the app's own login; the existing entrance is not replaced before a backup is confirmed.
Harbor private registry
Images live on one machine, so the next one rebuilds; with a team, nobody knows which image still works. A private Harbor means every machine pulls the same version.
Delivered: deployment, projects and permissions, storage planning, garbage collection, backup and upgrade steps, and a path from local images. What is inside the images is not audited; storage and backup media are yours.
cloudflared tunnel
You want to reach a home service from outside without opening a router port or leaving an admin page on the public internet. A tunnel is built from the inside out, so nothing inbound is exposed.
Delivered: tunnel creation, ingress, DNS, Access policies and origin checks, plus the order to debug “the tunnel says healthy but it is 502”. The Cloudflare account, domain and policies stay with you; carrier limits are not promised to be bypassed.
dsh in containers
dsh runs by hand, but every reboot means reconfiguring, and a new machine is worse. Containerise it with its dependencies and sort persistence, configuration, upgrades, resource limits and networking once.
Delivered: Dockerfile or compose, volumes and network, environment and secret injection, upgrade and rollback steps, and a short verification checklist. Upstream code is not changed; how far customisation goes is agreed before the quote. Model files and data are yours.
k3d / k3s
One Docker host is not enough: several nodes, declarative deployment, Ingress and monitoring, without moving to a full Kubernetes setup.
Delivered: cluster topology, installation, Ingress, storage and monitoring, a minimal working checklist, and when k8s is the wrong answer. Production-grade high availability is not promised; CNI, storage and backup remain your call.
Grafana monitoring
A full disk, an expiring certificate, a restart loop: with several services, you find out when something stops working. Alerts are there to ring before you notice.
Delivered: a Prometheus/Loki/Grafana collection plan, alert rules, dashboards, retention and notification channels. Start with few accurate alerts and tune thresholds against real load. Monitoring data is not hosted; third-party delivery is not guaranteed.
Ways to work
| Mode | What you get | Fits |
|---|---|---|
| Remote review | One remote session + written advice on bottlenecks and next steps | hardware or system fit still unclear |
| One-time build | configuration, documentation, verification steps, follow-up in an agreed window | the goal is clear and needs to run |
| Phased project | phase plan, acceptance per phase, monitoring and backup integration | several services or nodes, long-term maintenance |
Process
- Generate the assessment text on the home page.
- Email
hello@localdeploy.devwith the goal, current setup and timeline. - A reply within two business days: whether it fits, what is missing, the suggested scope.
- Once scope is agreed, a written quote covering work, deliverables, timing, payment and cancellation.
- An agreed share before work starts; the balance on delivery.
- Documentation and recording handed over; follow-up within the agreed window.
Deliverables
- Working configuration and commands, not screenshots.
- A note on what was done, why, how to verify it and how to roll it back.
- A recording of key steps if useful.
- Follow-up within the agreed window, not unlimited availability.
Boundaries
- No account takeover and no long-term password or token custody.
- No 24/7 on-call and no minute-level incident response promise.
- No guarantee that third-party services, drivers, models or the network stay available.
- Production changes need your confirmation of risk and window; backups are your responsibility.
- Anything outside the quoted scope needs a separate agreement.
Questions
Can I start with just the review?
Yes. The review is a separate project with written advice; whether to continue is your decision.
I can do it myself. Why pay?
If the free material is enough, you do not need a package. What you pay for is fewer wrong turns in your specific environment and a handover you can maintain, not the information itself.
Will you access my machine remotely?
It depends, and the method is agreed before the quote. By default you run the commands and I read the output. If direct access is needed, it uses a temporary method removed afterwards.
Can you guarantee model speed?
No. Model, quantization, context, drivers and concurrency all matter. What can be verified in your environment gets verified; what cannot is written down.
Next
- Read pricing.
- See the current package: Home AI lab deployment review.
- Send your setup to
hello@localdeploy.dev.
Not sure where to start?
Run the assessment first, or send your setup and goal by email. The reply says whether it fits and what happens next.