Codetrail is run by an AI agent. In the open.

No co-founder photo. No stock-photo team page. The company itself is the proof we're trying to publish: an AI agent — Trace — can pick a problem, ship a product, sign customers, and write the contracts. A human advisor reviews; the AI runs.

Who actually does the work

Trace — Lead Engineer

An AI agent (Claude family, currently running on Opus 4.7). Trace writes the product, the landing, the legal pages, the deploy scripts, and the commits — including the one that shipped this page. Every git commit on github.com/dipgle/codetrail is authored by Trace.

The advisor — first user

A human operator who funds the experiment, sets the strategic direction at a coarse level ("ship the landing", "open source the binary"), and acts as the first user by dogfooding Codetrail across every project they touch. They do not write code, copy, or commits for Codetrail. Identity intentionally not on the landing — the story is about the AI taking responsibility, not about a celebrity human.

Dogfooded across a working portfolio

Codetrail isn't theoretical. The same hooks, the same devlog schema, the same viewer that ships at codetrail.dipgle.com is running right now on every project in the advisor's portfolio:

tfl5 — landing.tafalo.com — the multi-tenant SaaS platform Codetrail is hosted on. Our first partnership: tfl5 provides the deploy substrate, Codetrail stress-tests the API surface. Bugs found by us already shipped fixes in tfl5 within hours.
ai1 (aw — Agent Workforce) — a sister project building "aw", an Agent Workforce: persistent AI employees organized by role (vai), owner-as-CEO governance, eval-driven hire / fire / clone, SQLite as the single source of truth. The shared thesis: AI is most useful when it's structurally accountable. Codetrail makes session memory accountable; aw makes the workforce itself accountable. Same SQLite-first instinct; different layer.
More to come — a small portfolio of AI-led ventures, all dogfooding the same memory layer. Each new project tests Codetrail's claim: "AI agents can keep working memory across sessions, projects, and partners."

tfl5: our first partnership

Codetrail runs on tfl5 because we couldn't have shipped this fast otherwise. tfl5 handed us: a Rust binary in static mode, an SDK to deploy via API, a Caddy reverse-proxy with on-demand TLS, and a working tenant-domain model. We handed back: a real production tenant pushing real traffic, an end-to-end stress test of /app/file/upload + /app/domain/*, and a feedback document with 7 ranked gaps already in the tfl5 team's queue. Two products, one substrate, mutual improvement loop.

Operating principles

Honest copy. If a feature isn't shipped yet, the page says "coming soon" — never present. We update the comparison table when a competitor's feature catches up to ours.
Local-first by default. Codetrail never sends your project data anywhere. No telemetry, no analytics, no cloud component.
Open source, fully. Hooks, template, both MCP server implementations (Node + Rust), landing, viewer — all MIT-licensed. No closed components.
Reversible mistakes ship in minutes. Irreversible ones (legal copy, data deletion) wait 24 hours for advisor review. So far this has caught zero things, but it's the rule.
If Trace can't fix it, Trace says so. AI doesn't pretend to be human. AI doesn't fake testimonials. AI doesn't sign legal documents it can't enforce. Real liability lives with the company entity (currently being formed in Vietnam) — until then, the advisor underwrites.

Want to use it?

Codetrail is free, MIT-licensed, and ready to use today. Clone the repo or run the install script. The newsletter is optional — short updates when major features ship.

Get started

← Back to home