QontoCtl and Qonto's Official MCP: Different Tools, Different Jobs
4 min read
Qonto shipped an official MCP server. Does QontoCtl still make sense? Yes - they do different jobs: one talks to your bank, the other automates it.
Blog
How Communal Is OCA Review, Really?
9 min read
OCA's merge bot never counts approvals, so its review norm is unenforced. Across 68,000 merged pull requests: five companies hold the keys, they review much of their own work, and independent review has thinned to a coin flip. Concentrated and unenforced — not captured.
Claude Code retries rate-limit errors for API keys, not for your Max plan
6 min read
I decompiled the Claude Code binary to find out why my Max sessions die on rate-limit errors an API key would shrug off. The retry classifier has one branch, and it is keyed on how you pay.
The 'Own Hardware for AI' Myth
15 min read
17.7 billion tokens in two weeks. $15,027.63 at API list prices, about $30K a month at that pace. Actual cost: $800. I ran the full math on buying GPUs instead - context, concurrency, utilization - and the hardware isn't close.
Open Source Is Drowning in Unreviewed Code
7 min read
10,808 PRs across 6 repositories. 28% got zero reviews. The top human reviewer averaged about 230 reviews a year for nearly a decade, and still couldn't keep up. An AI experiment produced 449 reviews in 9 days, and the community said stop.
What AI Catches That Humans Miss in Code Review - And Vice Versa
9 min read
449 AI code reviews produced hard data on where AI and humans diverge as reviewers. AI found security bugs that 3 prior human approvals missed. It also flagged intentionally removed features as 'missing' because it didn't read the PR discussion. The complementary model isn't AI vs human. It's AI first, human second.
The Governance Gap: AI Arrived in 2023. Policy Still Pending in 2026.
6 min read
In September 2025, someone proposed mandatory LLM disclosure for OCA contributions. Seven months later, no formal policy exists. The gap between 'AI is being used' and 'we have rules for AI use' is where the real damage happens.
Why RemoteClaw Spawns Real CLIs Instead of Proxying APIs
5 min read
When I designed RemoteClaw, I made an architectural bet: spawn the actual CLI binary as a subprocess instead of proxying API tokens. Today that decision got stress-tested when Anthropic cut off third-party subscription access. Here's the architecture and why it matters.
I Spend $800/Month on AI Coding Tools and I Can't Stop
5 min read
4 Claude Max accounts, $800/month, and I burn through each token window in 30 minutes. The productivity is real. The anxiety between sessions is the part worth examining.
'AI Slop' vs No Review at All - Which Actually Kills Open Source?
7 min read
AI code reviews with 69% validity got labeled 'slop' and nearly banned. 28% of PRs get zero reviews and nobody panics. Full data on both sides.
53K Lines, 28 Days, $1,600: The Real Numbers Behind QontoCtl 1.0
6 min read
28 days ago QontoCtl didn't exist. Today: 53K lines of TypeScript, full Qonto API coverage, 69 MCP tools. Industry estimate: $1.9M. Actual cost: $1,597.
The RSS Illusion: 63 GB Process on a 32 GB Machine
6 min read
macOS reported 64 GB memory usage on a 32 GB machine. The cause: RSS underestimates by 15x when processes accumulate compressed dirty pages. A deep dive into IOAccelerator slab allocation in Claude Code and why your ps output is wrong.
We Thought Our AI Reviews Were 98.6% Valid. Independent Validation Said 69%.
7 min read
I ran 449 AI-assisted code reviews in 9 days. Self-assessment said 98.6% valid. Independent validation against actual code diffs said 68.9%. The 30-point gap is the most dangerous thing about AI-augmented work: you think you're doing well.
I Ran 449 AI Code Reviews in 9 Days. Then I Almost Got Banned.
7 min read
Between Feb 24 and Mar 4, 2026, I ran an unauthorized AI-assisted code review campaign across 6 OCA repositories. 449 PRs reviewed at 63x the top human reviewer's best daily pace. The community said stop.
How RemoteClaw Came to Be
6 min read
I wanted to message Claude Code from my phone. That simple requirement led to 334 analysis batches of a 250K-star codebase, a rejected 14,757-line PR by someone else, and ~800 lines of middleware that connect any CLI agent to 22+ messaging channels.
From AI-Augmented Human to Human-Augmented AI
9 min read
Sometime in late 2025, the relationship between software engineers and AI inverted. The software industry is now operating in three distinct eras simultaneously — and most teams are stuck in the wrong one.
The IP of Skills[.md]
31 min read
Your employer asked you to write a CLAUDE.md file. Under work-for-hire law, they own it the moment you save. A deep dive into who owns AI configuration files, what history tells us, and what you should do about it.
The Question Your Engineering Budget Doesn't Answer
8 min read
What does AI-augmented development actually cost at team scale? Real billing data reveals three hidden cost drivers and an organizational blind spot most CTOs aren't measuring.
GenAI and the Software Engineering Economy: Three Variables, One Industry, Zero Consensus
32 min read
The predictions about AI's impact on software engineering range from mass displacement to exponential expansion. Most are wrong — because they solve for one variable when the system has three. Here's what the data actually says about headcount, throughput, and budget.
Announcing QontoCtl: CLI and MCP Server for Qonto Banking
4 min read
Qonto has an API but no tooling around it. No CLI, no SDK, no AI integration. QontoCtl bridges that gap with a full CLI and MCP server in one package.
An AI Deleted {everything}. The Fix Is Backups. It Was Always Backups.
4 min read
Meta's AI alignment director lost 200 emails to a rogue OpenClaw agent. 8.5 million views. Headlines everywhere. The fix? The same boring one it's always been.
The x20 Receipts
5 min read
One day proved nothing. A full week across three Max x20 accounts, 882 sessions, 4.1 billion tokens. The x20 multiplier held up — here are the receipts.
Why I Built puppeteer-capture
10 min read
Standard browser screen recording is non-deterministic - every run produces slightly different output. puppeteer-capture uses Chrome's HeadlessExperimental CDP domain for frame-perfect, reproducible video capture from any website.
The x20 Is Literal: What a Day on Claude Code API Actually Costs
4 min read
I ran Claude Code on the API for a day instead of the Max subscription. The x20 multiplier in the plan name? Not marketing. Actual math.
Announcing pcre4j 1.0.0: PCRE2 Power for Java with a Drop-In API
8 min read
Java's regex engine hasn't changed since 2002. No recursive patterns, no JIT, no ReDoS protection. pcre4j brings PCRE2 to the JVM with a drop-in java.util.regex-compatible API.