local·first·lab
~/lab · lanewatch --demo · synthetic demo data
lanewatch rendering local inference servers and cloud assistant lanes in one frame: braille pulse graphs per lane, a spend burn bar, and a rolling activity log
Local-first software and supervised AI, tested on real hardware.

Local First Lab

Hands-on comparisons of privacy-respecting, self-hostable software: the open-source apps that replace surveillance SaaS and keep your data on your own hardware.

New flagship: lanewatch, a btop-style terminal monitor for local inference servers and the cloud assistant lanes running on your machine. Python stdlib only; readings are numbers, never content. The story →

Security engagement: Hardening Loop, a security engagement that ends at proof each finding is fixed, not a PDF that gets filed. See how the loop works →

I also build these systems for clients: see the supervised-ops demo.

lab log · dated results, sources linked
2026-07-25 · lanewatch security pass
a hostile endpoint could paint 2,000 cells with 8 bytes
or drive memory to 1.3 GB with one 137 MB reply.
both fixed in code and pinned by 25 security tests.
2026-07-17 · task-success bench, 14 browser goals
hand-written, ceiling    │ 14/14
36B local model, on-site │ 3/14 (baseline 0/14)
4B local model, on-site  │ 2/14 (baseline 0/14)
authoring validity       │ 14/14
wire format, not model scale, drove the dominant failure.
full data: lab results on GitHub
type bench in the terminal to watch a run
~/lab · replace.sh · a real prompt: type a command, try help or projects
you@localfirstlab:~$ replace "AdBlock Plus"
 adguard ● self-hosted: your data stays home
you@localfirstlab:~$ 
Start here

What this lab actually ships

lanewatch
A btop-style monitor for local inference servers and cloud assistant lanes. Readings are numbers, never content; Python stdlib only.
lfl-terminal
A deterministic browser terminal with a supervised local-AI lane: every action approved by you, nothing sent to a cloud.
lfl-lab
An open research bench that stress-tests lfl-terminal's trust boundary with adversarial pages and measures what local models can really do.
supervised-ops-demo
A working demo of supervised AI operations: the model drafts, a human approves every outgoing action.

Lab log: 2026-07-25: adversarial security pass on lanewatch, proving a hostile local endpoint could repaint the terminal or balloon memory to 1.3 GB, then fixing both and pinning 25 security tests. Before that, 2026-07-17: task-success bench, a 4B and a 36B local model across two conditions (2x2) over 14 scripted browser goals: authoring validity 14/14 everywhere, task success 0-3/14. Wire format, not model scale, drove the dominant failure. Full results →

0 trackers0 cookies0 third-party requests71 tools tracked20 comparisons
Live (computed on your device, by your browser)

What your browser just told us

Any site you visit can read a stack of signals the instant the page loads: no cookies, no permission prompt. This page is no exception. The difference is what happens next.

Consulting

Supervised AI systems for small businesses

I design and install AI operations systems where a human approves every outgoing action (inquiry triage, drafted replies, follow-ups, audit trails). Fixed-price pilots from $500. Try the interactive demo.

Try the interactive demo → Or email hello@localfirstlab.org
Interactive: runs entirely in your browser, nothing sent anywhere

What are you overpaying for?

Check the subscriptions you already pay for. We'll tally the annual cost and show the self-hosted alternative already in our catalog, right next to it. Prices are 2026 estimates for common individual/family plans; yours may differ.

You're paying $0/yr
Self-hosted cost $0/yr
Estimated savings $0/yr

Honesty note: self-hosting isn't free. It costs your own time (setup + upkeep) and hardware/electricity: a different bill, not a $0 one. Figures above are estimates for the subscription side; your mileage on both sides will vary.

Want it set up without the headache? See how we do this for people → or email hello@localfirstlab.org.

Also from this lab

lfl-terminal: a terminal for your browser

A Chrome extension that drops a command line onto any site: known commands run deterministically, anything else becomes one action a local model proposes and you approve before it touches the page. Open core, Apache-2.0, early (v0.5.x, Chrome only, load-unpacked for now).

See it in action → GitHub →

Compare all 71 alternatives in one table →

Need this set up for you?

I do self-hosting migrations, privacy stack audits, and private/local AI installs, including for confidentiality-bound firms. About me · GitHub · Get in touch

localfirstlab.org · home · theme dark · crt low · 0 trackers · 0 third-party requests