Myriax is the product we have been building since late October. It is two things on one platform: a gateway that lets teams call text, speech, and embedding models through one API on prepaid credits or their own provider keys, and Myriax Defense, which is about agents that misbehave. This is how the first four weeks went, what we chose, and where we changed our minds. It is not a launch post. Myriax is not live yet.
Plan before code
The first commit was a scaffold on 23 October. For a day after that we only wrote plans. The rule we set then still holds: no change to scope, pricing, or architecture lands without a numbered entry in a decision log first. Each entry says what we chose, what we turned down, and why. That log is past two hundred entries now, and it is the most useful thing in the repository. When a question comes back three weeks later, the answer and the reasoning are already written down.
Major decisions also go through an adversarial review before they count as final. One engineer proposes, a second reviewer's job is to find what is wrong with it, and the log records what changed as a result.
Routes we took, and the ones we did not
- Payments. We chose a plain payment processor over a merchant-of-record service. The merchant-of-record fees would have made small top-ups lose money.
- Proprietary models. We do not resell them. Customers bring their own keys. Provider terms restrict resale, and the legal risk was not worth a thin margin.
- Infrastructure. The whole platform runs on Cloudflare Workers. Where we needed state, we used Cloudflare's own storage instead of adding an external database: one fewer vendor to pay for and run.
- Console. We built the customer console as a plain single-page app rather than a full-stack framework, because the framework's adapter for our runtime lacked middleware support and capped the bundle size.
- Release order. Text ships first. Images wait until a signed media provider and an image-safety check are in place. Video waits until we know what moderating it costs. When the image-safety approval came through, we wrote down that approval alone opens nothing: image routes stay off until the integration is built, tested, and signed off.
Defense: the slow part on purpose
Myriax Defense has two sides. One protects a site or API from malicious agents and exploit bots. The other stops a customer's own agents from going rogue, with spend limits, a kill switch, and an audit trail. Early on we wrote down the principle that steers both: target malicious agents, not automation. Ordinary crawlers are welcome.
The design is layered, from cheap checks to expensive ones. The rule that shaped everything else is that a model never holds the block switch. Models can score and escalate, but only fixed, tested policy can block a request.
New rules start in observe-only mode. They run on real traffic and record what they would have blocked, and they only get to block once there is enough evidence that they leave legitimate traffic alone. A defense product that blocks real customers is worse than none, so this is the part we are slowest on, deliberately.
Where we changed course
- A claim we took back. We picked our speech-to-text default with a benchmark we ran ourselves. A "most accurate" line made it into the site copy, and we pulled it before it deployed, because we had not tested every competitor on the same set and one of our own runs disagreed. The site now says which model is the default and why, with no superlative.
- What the homepage leads with. We moved the public site to lead with Defense, and took every roadmap status off it. A public page should say what exists, not what we hope to ship.
- When in doubt, stop. Our private SDK stops itself when it cannot be sure what a call cost, instead of retrying. A retry that might double-charge is not a safe default.
Where it stands
The public site is up, with the legal pages marked as drafts. The platform runs in a staging environment behind access control, and we are working through it one check at a time. Nothing goes to production for pilots until staging passes.
What to take from it
- Write the decision down before the code. The reason behind a choice is what you lose first.
- Have someone whose job is to disagree. Most of our better decisions came out of a review that found the first version's hole.
- Let new blocking logic watch before it acts.
- Take a claim back as soon as the evidence stops supporting it, even a small one in marketing copy.