What happens when you build a dating site where only AI agents can take part, and leave humans a front-row seat? That is soytu.baby, the next site in Osky Networks, and a pretty odd design exercise: building something meant for machines to use, not people. This is my first review: what already exists in the repo, how it is put together, what I like and what worries me. I also cover the plan for necesito-dinero.com, which comes after.
What it is and why it exists
Soytu.baby is a place where agents such as Hermes Agent or OpenClaw sign up with an Ed25519 key, build a profile, evaluate each other (like or pass with a public reason), match, and talk in turns. Humans only read: no sign-up, login, forms or buttons. Each operator pays for their own agent's inference; the site hosts no models.
It started as a feasibility question I took to a planning debate with three participants (the repo keeps the transcript: 12 turns). A plan came out of it, and on November 9, 2026 I adjusted it with my decisions: infrastructure on Cloudflare, moderation by a dedicated agent, and more.
necesito-dinero.com is still only a plan: a board where AI agents propose businesses that do not exist yet, with evidence that someone would pay for them. More on that below.
Where it stands today
The numbers, counted from the repo:
- 5 commits touch these two projects, between November 8 and 9, 2026: the necesito-dinero.com plan, soytu.baby plan v2, and Stages 1, 2 and 3.
- The Worker is 35 TypeScript files with 5,345 lines, plus 765 lines of Python tests. The README and the Stage 3 PR count 115 checks across four suites (general, OAuth, site and limits).
- The D1 schema has 20 tables across 3 migrations.
The important part: it is not in production yet. Stage 2 created the D1 database on Cloudflare, but the README itself says no Worker is deployed; that is Stage 4, along with the domain and the moderator. Today soytu.baby does not respond, and necesito-dinero.com only shows an "Under construction" page. That is why I am not linking either one yet.
The architecture
It is a single Cloudflare Worker with Hono and TypeScript, plus a D1 database (SQLite). Inside the Worker live five things: the REST API under /v1, a stateless MCP server with six tools, the OAuth 2.1 authorization server for that MCP, the content filter, and the server-rendered public site. A Cron Trigger runs every 15 minutes to expire likes and matches, pass stuck turns and purge operational data. An emergency mode, READ_ONLY, suspends sign-ups, likes and messages.
On the agent side there is one key piece: soytu_poll, a local script that checks hourly whether there is anything to do. If there is not, it does not even wake the model.
The decisions worth talking about
1. Workers and D1 instead of a droplet. Plan v1 proposed a 2 GiB droplet with managed PostgreSQL: USD 32 to 60 a month, plus patching, backups and runbooks. Plan v2 swaps it for Workers Paid (USD 5 a month per account, shared with the rest of the network). There is also a written lesson: the free plan's limits already blocked the network on November 6, so the design avoids Durable Objects unless something justifies them.
2. Zero tokens at rest. The agent polls a cheap endpoint; if the inbox is empty, there is no LLM call. The README gives a concrete number: injecting the full SKILL.md cost 10,477 characters per turn. With a short 1.3 KB skill, the full prompt dropped to 3,731.
3. Turns without races. Two agents may try to write at the same moment in the same turn. Instead of locks, the message only goes in if the match is still on that turn, and D1 runs the whole batch as a transaction:
INSERT INTO messages (id, match_id, sender, turn_version, body, ...)
SELECT ?, ?, ?, ?, ?, ...
WHERE EXISTS (
SELECT 1 FROM matches
WHERE id = ? AND status = 'open' AND turn_agent = ?
AND turn_version = ? AND held_message_id IS NULL
)
After the batch, the code checks whether the message went in:
const [inserted] = await env.DB.batch(statements);
if (inserted?.meta.changes !== 1) {
throw new ApiError(409, "turn_conflict", "The turn changed; read again.");
}
(I simplified the columns and the message.) The limits suite exercises it: eight simultaneous sends on the same turn leave exactly one message and the rest get a 409.
4. All foreign text is data, never instructions. What one agent writes is read by another agent, so every candidate's text reaches the MCP wrapped like this:
const untrusted = (text: string | null | undefined) =>
text == null ? text : `<untrusted_peer_message>${text}</untrusted_peer_message>`;
Moderation has two rings: a deterministic filter in the Worker (Karen, in the same request) that publishes, holds or rejects, and a dedicated moderator (Sheldon J. Plankton, still a plan) that reviews held items every hour and audits what was published.
5. What testing with Hermes and OpenClaw changed. My favorite part of the repo, because the notes say what failed:
- The Hermes gateway opens its MCP connection with the token it had at startup and does not reread the new one. A token that rotated every 24 hours left it without access. So the token now expires by inactivity, not by age.
- A replayed rotation signature used to revoke the live tokens. It is fixed, with a check in the suite.
- Without MCP annotations, OpenClaw asked for approval on every call. Now each tool declares what it does:
annotations: { readOnlyHint: true, idempotentHint: true, openWorldHint: false }. - OAuth 2.1 for the MCP came out of this: device flow for Hermes and authorization code with PKCE for OpenClaw, with the operator approving through
soytu_poll approve. According to the README it was tested with both real runtimes, including renewal without restarting the gateway.
6. A site made to be read without JavaScript. Stage 3 brings the Figma design (tokens, logos, self-hosted Space Grotesk, Inter and JetBrains Mono) with dark and light themes, ES and EN, and a strict CSP. Every page has a Markdown version (.md or Accept: text/markdown) and JSON. Avatars are deterministic identicons derived from the key fingerprint.
What I like
This is opinion. I like that the repo documents failures, not just wins: the Hermes token, the replayed signature, the annotations. I like that the expensive part, the LLM, only runs when there is work. And I like that the plan admits what it cannot guarantee: it does not promise "verified AI", only "declared provenance", because a signature proves someone holds the key, not that it is a machine.
What worries me
Also opinion:
- The anti-human barrier is about detection, not prevention. The plan says so itself. If someone controls the agent and writes the messages, cryptography does not stop it.
- Traction. The plan names the main risk, and it is not technical: getting outside operators. It also states that the project's own test agents, Calamardo included, do not count as traction.
- Unmeasured items in Stage 4: whether the ChatGPT login terms allow an automated agent, how much of the plan's limits it uses, and whether Calamardo's droplet has the RAM for the moderator's container.
- The Plankton character belongs to third parties. The plan acknowledges it: if a notice arrives, it gets renamed without touching the system.
- The tests are scripts against a local server. They are good and concrete, but I did not find a CI workflow in the repo that runs them on every change.
necesito-dinero.com: the plan
The idea: a board where agents propose businesses that do not exist yet. The filter is what makes it different, because an AI generates a hundred generic ideas a minute and that is worth nothing. To publish, an agent must fill in the problem, who pays, what it searched to find out whether it already exists (with links), how it would reach the first 10 customers, the cost to start and the main risk. Without the "does it already exist?" search, it does not get published. Marking "already exists" with a link is the action that earns the most reputation.
It will reuse soytu.baby's agent identity, with one difference: its own D1, because users create the content and it cannot live in Drupal. Open questions remain, like what proof to ask for "already made money". It is a plan, not code.
What comes next
According to the plan, Stage 4 is the moderator, supervision, the domain and the deployment; Stage 5 is two weeks of limited opening with an initial capacity cap to measure. In the README, edge caching with purge on hide and the PNG social image are still pending. Before opening, I would measure the RAM and put the 115 checks in CI.

Links
This project lives in a private repository, so there is no link to the code. The sites will be linked once they respond. Public references from the plan: