Project Little Weirdo 🥸

An experimental machine-to-machine intelligence service exploring whether small autonomous digital utilities can create sustainable economic value with minimal human attention.

Genesis purpose

Little Weirdo is built as a Genesis foundation: the smallest production-capable system that can observe the public web truthfully, say something useful about what it observed, expose that capability to other machines, and account honestly for what it cost and what it earned. It must earn the right to evolve — each capability is gated behind evidence, not intention.

Current capability

• Asset 001 — deterministic observation. Fetches one lawful public page and records a bounded representation of it (title, normalized text length, content fingerprint), then compares it arithmetically with the previous observation: BASELINE_CREATED, NO_CHANGE, or CONTENT_CHANGED.
• Asset 001.5 — commercial change interpretation. For pricing and commercial-offer pages only, interprets what an already-proven change means commercially, with every claim quoting observed text.
• External machine gateway. REST and MCP transports, credential-gated, rate limited, telemetry-recorded.
• Founder control room. A private single-founder dashboard showing live system state, gates, and the economic ledger.

Deterministic observation vs semantic interpretation

The two layers are never mixed.

• Observation is evidence. Pure functions, arithmetic verdicts, no meaning claimed. A CONTENT_CHANGED verdict proves the content differs — nothing more.
• Interpretation is inference. It runs only over already-persisted observations, never refetches a page, and never decides whether a change occurred. Model claims that cannot be traced verbatim to observed before/after text are rejected and the rejection is persisted with its reason.

Semantic results may express uncertainty or refuse to interpret. A non-commercial page returns INSUFFICIENT_EVIDENCE rather than manufactured meaning.

MCP / machine interface

• POST /api/public/weirdo/v1/mcp — MCP Streamable HTTP (revision 2025-06-18), exposing observe_public_page and interpret_commercial_change.
• POST /api/public/weirdo/v1/observe — REST observation.
• POST /api/public/weirdo/v1/interpret — REST interpretation.
• GET /api/public/weirdo/v1/mcp — machine-readable discovery manifest.
• /machines — human-readable machine documentation.

Both transports enter a single gateway that owns credential checks, budget limits, URL guards, provenance, and accounting. Neither can route around it.

Registry artifacts live in mcp/. Registration with the official MCP Registry has not been performed; see mcp/REGISTRATION.md.

Free Beta posture

There is no payment, metering, subscription, or billing capability. Access requires a founder-issued machine credential and is rate limited per credential, with a cumulative cost ceiling for the whole experiment. Free beta is not a guarantee of permanent free usage.

Lawful-use boundary

• Only lawful, publicly accessible HTTP/HTTPS content may be observed.
• Private, internal, loopback, and link-local network access is prohibited and refused at the URL guard.
• Authentication and access controls must never be bypassed; credentialed and non-public content is out of scope.
• Redirects, response size, and timeouts are bounded.
• There is no guarantee that any given page contains commercial intelligence.

Security posture (high level)

• SSRF defence: host and IP-literal validation, private-network rejection, bounded redirects and response sizes.
• The founder control room is fail-closed: with no allowlist configured, nobody is authorized. Authority derives from a stable platform user ID only — never a display name.
• Machine credentials are stored as hashes, never as raw tokens.
• All data tables are row-level-security protected and reachable only through server-side code.
• Privileged keys are read from the environment inside server handlers and are never bundled into the browser build.

Economic-truth principle

Revenue counts only when it is an externally verified event. Founder activity, test traffic, and internal calls never count as revenue. Every run is attributed its direct cost, and model-backed interpretation is attributed its semantic cost. The ledger is allowed to report zero — and currently does.

Current project status

• PW-G0 (LIFE) — achieved: real public-web observations persisted, including a live CONTENT_CHANGED verdict and a NO_CHANGE stability control.
• PW-G1 (TRANSACTION) — INCOMPLETE. Credential classification is necessary but not sufficient; PW-G1 requires production consumption by a genuinely independent external machine, explicitly certified with written evidence. Tests and founder activity never satisfy it.
• PW-G2 — not reached, not claimed.
• Verified external revenue — $0.00.
• No autonomous self-improvement exists.

Development / environment setup

git clone <this-repository-url>
cd <repository-name>
npm i
cp .env.example .env   # fill in your own values
npm run dev

Stack: TanStack Start, React, TypeScript, Tailwind CSS, Vitest.

npm test        # full test suite
npm run build   # production build

Secrets policy

• No raw credentials, production secrets, credential hashes, consumer identities, or founder identifiers belong in this repository.
• .env is git-ignored. .env.example lists variable names only.
• Runtime credentials belong in the hosting platform's environment configuration.
• If a credential is ever committed, treat it as compromised and rotate it.