HOW THIS SITE WORKS
This site is a working demonstration. Cecil — the AI secretary on the landing page — is a prompt-only agent: one language model, one carefully written brief, and a set of limits enforced in code. If you are evaluating Vinga's work, the machinery behind this page is part of what is on display.
The agent
Cecil runs on a single system prompt. Everything he knows about Vinga sits in a public dossier inside that prompt — there is nothing hidden behind it. A Claude model generates his replies; a Gemini model summarizes older turns so long conversations stay coherent. Conversations live in server memory only: they are not written to disk and do not survive a restart.
The budget
A public AI endpoint can be drained by anyone with a script, so Cecil's day runs on a fixed budget — spend is computed from real API usage on every turn, per conversation and globally. When the day's budget is spent, he retires until tomorrow and the leave-word form takes over. Conversations are budgeted, and Cecil says so plainly when a limit is reached — bounded operating cost is part of the design, not an apology.
Messages
Word reaches Vinga two ways: the leave-word form, and Cecil himself — ask him to pass a message on, and he will. Both write into the same store: a database on a persistent volume holding contact, message, and timestamp. The row holds what you chose to send and when — no IP address, no fingerprint. Deleting your message is deleting one row — write to vinga@skyforge.sh and it is gone.
The two tools
Cecil has two capabilities beyond conversation. The first is recording a message for Vinga. It is write-only — he can send word out, never read anything back — and consent-gated: nothing is recorded unless you ask or agree. If he offers to attach a few lines of his own context to your message, that is separately yours to decline, and he tells you what they will say before anything is sent. The limits are enforced in code, not by asking him nicely: two messages per conversation, daily caps on storage and email.
The archives clerk
The second is a research agent Cecil can send into one of Vinga's public repositories. With your consent to the wait — it is a genuine exploration and takes a few minutes — the clerk downloads a fresh copy of the repository at its newest commit, reads it with read-only tools, and returns a report. Cecil relays it with receipts: the file paths become links pinned to the exact commit consulted, so every claim can be checked at its source. An honest "that is not in the repository" counts as a good answer here. One trip per conversation, daily caps in code, and the clerk reads only what anyone could read on GitHub.
What is recorded
When Cecil cannot answer a question about Vinga from his dossier, the question is logged — the question alone, with no identity attached — so the dossier learns what visitors actually need. No conversation transcripts, no visitor profiles.
Visits are counted, too. A site that exists to reach people should be able to say whether it does — so each page request is tallied into a daily, aggregate count: the date, the referring domain (where the visit came from), a rough browser-or-bot label, and a country. That is the whole of it. No cookies, no fingerprinting, no third-party analytics script, and no per-visitor record — the count has nowhere to store who you are. We measure our own traffic, first-party and in the open. The country is resolved from a local copy of DB-IP's IP-to-Country Lite database (licensed CC BY 4.0): your IP address is read once to look up the country and then discarded, never written down. No cookies are set at all until you start talking to Cecil.
Testing
Changes to Cecil's brief ship through a behavioral regression suite: scripted visitors — the awkward ones included, from rate probes to injection attempts — run against the real system before anything reaches this page.
Questions about any of this: vinga@skyforge.sh