Short answer: we use Claude Code, Anthropic's coding agent, every day: under written project rules, under hooks that block unsafe commands before they run, and with a human review before anything is merged or deployed. An AI agent may read your systems; anything that writes, spends or sends waits for a person. Our articles are drafted with Claude from our own records, and our founder checks every claim and approves every piece. Each article, this one included, says so in its first line.
Where do we use AI?
In three places, and our How we use AI page states each one as a rule:
- In our writing. Articles on this site are drafted with Claude from our own records: code, configuration, measurements and public sources, each listed at the end of the article.
- In our engineering. Claude Code works in our repositories every day, under the rules and checks described below.
- In what we build for you. Agents we build keep their API keys on your server and work only through tools we can test. Any action that changes your data needs approval. How an agent keeps its keys and how an MCP server keeps writes behind approval show the mechanics.
How is an article like this one made?
Every piece follows the same six steps:
- A brief that lists every claim and its source: a repository, a configuration file, a public page or a dated record.
- A draft in English and Spanish by Claude. The Spanish is written for a Colombian reader, not translated line by line.
- A fact-check pass that re-reads each claim against its source and lists what it could not verify. A claim with no source is cut, not softened.
- Our founder's decision: publish or not.
- The deploy. Articles live as files in the site's repository, so each one ships as a reviewed change through the same tests as the code.
- A look back at search and AI-assistant referrals two weeks later.
Steps 1 to 3 and 5 to 6 are done with Claude. Step 4 is a person's, and only a person's. This article went through those steps. The agent that drafted it worked from our repository and our dated records, and it was not allowed to load our live sites at all.
Why it matters to you: every fact on this page has a source you could ask us for, and a person answers for each one.
How does the agent work on code?
- Written rules, per project. Each repository carries instructions the agent reads before it starts: the conventions, the commands, and what is off limits.
- Hooks that block, not advise. A hook is a check that runs before the agent's command does, and refuses it. One example: since 23 September 2026, a hook has refused a class of shell commands that had silently corrupted files on our workstation. The rule against those commands had been written down for weeks and was still broken. Once it ran as a hook, it held. When a written rule keeps failing, we turn it into a check that runs every time.
- The same gates as anyone's code. Every push runs the tests on a fresh install and audits the dependencies for known security advisories, and a deploy is not done until a health check on the live site passes. Rolling back means running the previous image. None of this cares who wrote the change.
- A person reviews and decides. A human review comes before anything is merged or deployed. On our own sites, a push to the main branch deploys, and that push starts only on our founder's word.
- The history says so. A commit written with Claude carries a co-author line. Between 1 September and 27 October 2026, 128 of the 165 commits on this site's repository did.
Why it matters to you: you get the speed of an agent and the controls of a careful team, and you can see in the history which changes the agent helped write.
Where don't we let it act?
These are the lines we don't cross, for our sites and for yours:
- Anything that writes, spends or sends waits for a person. An agent may read your systems. It does not change them, pay for anything or send a message without a person's yes.
- No deploy to your production site on its own. An agent can prepare a change. A person decides when it goes live.
- Your data stays out of a model unless you agree. We do not put your data into a model without your agreement.
- No invented clients, quotes or results. Nothing on this site is a synthetic person, a made-up testimonial or a number we can't show.
- No productivity percentage. We never measured a before-and-after baseline, so we don't quote one. We show the mechanism instead.
What does drupal.org ask of AI-assisted work?
Drupal's own "Policy on the use of AI when contributing to Drupal" was last updated on 20 October 2026. In short: if AI wrote a significant part of a contribution, you disclose it, however carefully you reviewed it. You are responsible for every word and line you post, and you must be able to explain it. Unreviewed AI output is not allowed, and comments are written in your own words, not pasted from a model.
We follow it there as here. Our founder's module elevenlabs_agent says in its own README that it is heavily AI-assisted and human-reviewed. Issue comments and patches that used AI say so, and are tested by a person before they are posted.
Why it matters to you: the Drupal community and we hold AI-assisted work to the same test, the same one you should hold any vendor to: disclose it, review it, and stand behind it.
Does your Drupal site need AI in it?
Not necessarily. On our own site the AI feature is the voice concierge on the home page, built on elevenlabs_agent (in its 1.0 alpha series), and nothing else on the site depends on it. As of 28 October 2026 we do not run Drupal's AI module on our own site. When AI belongs on a site, we build it under the same lines: keys on the server, tools we can test, and a person in front of every change.
What we don't promise
Today the engineer is our founder, and support runs in business hours, Colombian time (UTC-5), with no 24/7 on-call. The one fixed price is an hour of the founder's time: USD 75. Our How we use AI page states the rules above for every engagement.
To talk about AI on your Drupal site, or about how we would build it, see what we offer or write to us.