BareStack

BareStack We build the systems you don't want to.

19/06/2026

Most small-build quotes I see are off by a factor of three.

Either too low, because someone quoted the "happy path" and ignored the part where the client changes their mind twice. Or too high, because the freelancer padded it for the same fear.

Neither wins.

The fix I use for any project under 10k:

1. List the three things you are actually being paid to deliver. Not the work. The deliverable.
2. Add a line called "rework budget". For small clients it should be 20%. For bigger ones, 40%. If you do not need it, great. You just made 20% extra. If you do, you are not eating the cost.
3. Quote a range, not a number. "8 to 12k depending on X" beats "10k". The floor is the price you would walk away at. The ceiling is what it costs if they are a nightmare. Most jobs land somewhere in the middle.
4. Half up front. Half on delivery. No milestones on a two-week build. Milestones are for projects longer than a month, when scope drift is the real risk.

The boring math above is the difference between a profitable agency and a busy one that wonders why the bank balance never moves.

The most powerful AI in the world right now is the one you can download for free.Last week the top 4 most-used models on...
17/06/2026

The most powerful AI in the world right now is the one you can download for free.

Last week the top 4 most-used models on OpenRouter came from one country. Not the US.

Cheap to build. Free to run. No one can revoke it.

That's not a prediction. That's the leaderboard this morning.

barestack.org

15/06/2026

Microsoft just announced Scout. An AI agent inside Word, Outlook, Teams and SharePoint that watches what you do and starts doing tasks before you ask.

It runs on their servers. Reading your email, your files, your client work. On a monthly seat, forever.

The bit the launch skips: you do not need them for this.

Open-source agents already do this kind of work, on your own machine, with your data never leaving it. Hermes Agent is one. MIT licensed, self-hosted, and it will happily read your inbox, file your docs, chase invoices and run your tools. It is what we run our own ops on.

github.com/NousResearch/hermes-agent

Big tech wants you renting an agent that phones home. The open-source version is free, yours, and usually just hidden behind worse marketing. We help small businesses set these up so they own the thing instead of renting it.

12/06/2026

The cheapest piece of software I ever bought for my business was a sentence.

It was on the bottom of every invoice:
"Interest charged at 8% per month on balances over 30 days."

Most clients never read it. The 2% who paid late did. Because now it wasn't me being awkward on a follow-up call. It was a clause.

Three things I learned chasing invoices for 6 years:

1. The contract matters more than the chase. If your terms say "due in 14 days" you can ask on day 15 without flinching. If your terms say "net 30" you'll dither till day 45 and the client knows it.

2. Invoice the moment the work is done, not at the end of the month. Speed of invoice is the single biggest predictor of speed of payment. Batching feels tidy. It costs you 2 weeks of float.

3. The first follow up should be a question, not a statement. "Just checking this landed" gets a reply. "Please pay this" gets a defensive silence.

You don't need a collections agency. You need a system that runs without you feeling like a bad guy.

11/06/2026

I run a small software agency. Every week a small business owner asks me "do I need a CRM?"

Honest answer I keep giving:

No. You don't. Not yet.

You need a CRM when you have a problem a CRM solves. Not before.

If you have got 20 active customers, a spreadsheet and a phone is the right tool. If you have got 200 and you are losing track of follow-ups, then yes.

The trap I see:
Small business buys a CRM at 15 customers because the SaaS ads say they "need one." Now they spend 4 hours a week updating it. They hate it. They stop using it. Six months later they still pay for it because cancelling feels like admitting defeat.

That $19-99/month per seat isn't a CRM. It's a guilt subscription.

The rule I use for any new tool:
Don't buy it until the pain of NOT having it shows up in your week for the third time.

Spreadsheet, email, calendar, and a notebook. That stack gets a service business to $500k revenue. Most of the tools you think you need are solving for a business 10x your size.

If a tool doesn't earn its place this month, kill it. Subscriptions compound quietly.

Microsoft just unveiled their first in-house reasoning model. This week.The line they used: "lessen reliance on OpenAI."...
04/06/2026

Microsoft just unveiled their first in-house reasoning model. This week.

The line they used: "lessen reliance on OpenAI."

The line they didn't use out loud: stop renting, start owning.

The cloud-AI playbook is being rewritten. Not in a press release — in a developer console.

For a decade, the safe move was "let the vendor do it." That move just got more expensive.

Microsoft figured it out: build your own. Run it on your own hardware. Pay once.
barestack.org

The enterprise AI market doesn't know what to charge for it.Per token. Per task. Per seat. Per "outcome" — whatever that...
04/06/2026

The enterprise AI market doesn't know what to charge for it.

Per token. Per task. Per seat. Per "outcome" — whatever that means.

SAP's president said it out loud this week:
"Everyone is scrambling to justify their investments."

The market is trying to squeeze AI into old pricing shapes — compute, seats, subscriptions.

But AI isn't infrastructure. It's closer to labour.
And you don't pay someone per keystroke.

Meanwhile: JetBrains just open-sourced Mellum2. A fast coding model. Free. The kind of thing that makes enterprise pricing look absurd.

If your AI bill is a moving target, you're funding someone else's experiment.
Run it on your own hardware. Pay once.
barestack.org

A US company just released the largest open-weight AI model America has ever shipped.550 billion parameters.300+ tokens ...
03/06/2026

A US company just released the largest open-weight AI model America has ever shipped.

550 billion parameters.
300+ tokens per second.
About 5x faster than closed-source rivals. Roughly 30% cheaper to run.

That was Nvidia's Computex moment this week.

Then you read the second line.

A Chinese lab called Moonshot is still ahead. Their open model Kimi K2.6 scores 54 on the same benchmark. Nvidia's Nemotron 3 Ultra scores 48.

The frontier isn't OpenAI vs Anthropic anymore. It's open weights vs locked-down weights. And the open side is being written in Beijing.

The biggest AI labs in the US still lock the best stuff behind an API key and a monthly bill.

Open wins because it's usable. Locked-in wins because it's FOMO. We're about to find out which one scales.

If you're building on AI and you want to know what your bill looks like next month, that's why we build software you run yourself.

barestack.org.

Microsoft is at Build this week. The Verge: 'Trust in Windows and GitHub is at an all-time low.'That's the lede. Not the...
02/06/2026

Microsoft is at Build this week. The Verge: 'Trust in Windows and GitHub is at an all-time low.'

That's the lede. Not the new reasoning model. Not the Copilot super app.

Microsoft's response: a cleaner Windows 11, pre-installed tools, an apology in product form. Fine.

The real tell? They're going local-first with AI. RTX Spark. Local inference. 'Instead of relying on costly cloud models' — their words.

The cloud-AI era isn't ending with a bang. It's ending with a developer console.

—

BareStack has been self-hostable from day one. No credit card. No subscription. No 'we may use your data.'

barestack.org

Address

Manchester

Website

,

Alerts

Be the first to know and let us send you an email when BareStack posts news and promotions. Your email address will not be used for any other purpose, and you can unsubscribe at any time.

Shortcuts

Share

Category