live · Next.js · Postgres · Qdrant
Struvo
Voice-to-field-report platform for construction teams: talk through a site visit, get back a structured, photo-linked report.
Visit →2026
The trophy case · The shipping record
Proof of work.
Everything the Studio's operator has shipped: products in production, open-source tools you can clone tonight, and the lab experiments behind them. “You can just build things” is a claim. This page is the evidence.
35projects
21open repos
5products
3GitHub orgs
Products
Software with users on the other end.
in production · Next.js · Drizzle · Neon
TwentyFive
Campaign and financial CRM built for a working creative agency: pipelines, scheduling, lead scoring, real books.
2026
in production · TypeScript
Bree'ze Squeeze platform
Operations platform for a family-run juice business: orders, batches, markets.
2026
in production · TypeScript
Friendform
Forms builder for creating, managing, and scoring custom questionnaires.
2025
building · React · FastAPI
FutureTree
Scenario planning and decision intelligence: decision trees with probabilistic outcomes instead of gut-feel strategy.
2026
The next product usually starts as a Saturday demo.
See the lab ↓Open source
Clone it, run it, break it, tell us.
These 21 repositories contain code and instructions. Many are parts of the Studio's AI setup. Some compare answers from several models; others check code changes, provide reference material or help run everyday tasks. The tools and checks used around an AI agent are often called its harness. The Studio's setup is documented as entry 001 on the Harness Ledger.
Published as-is, from real work. To clone a repository is to copy its code to your computer; check its setup instructions and license before using it. If one saves you an afternoon, come say so at a Sharathon.
Start here
Lucface · TypeScriptSnapshot · 2026-07-09
blind-council
Five frontier models debate blind; a sixth judges, corrects, and settles. A multi-model council with a dedicated judge seat.
GitHub →pushed 4 days ago
SanDiegoAIStudio · PythonSnapshot · 2026-06-17
apollo-mission-control
Multi-agent simulation of Apollo-era Mission Control: every flight-controller seat is an AI agent, together flying a KSP Apollo 11 mission.
GitHub →★1 · pushed 4 days ago
SanDiegoAIStudioSnapshot · 2026-05-11
sharathon-01
Set up for the studio's first Sharathon so each attendee could add one file. It holds a README, a license and an empty attendees folder.
GitHub →★1 · pushed May 2026
01
Councils & agents
More than one model on the problem: debating blind, flying a mission, or running a human-gated Intelligent OS.
- dan-the-automator
Public Intelligent OS control plane: kill-switch, heartbeats, human-gated actions, and a living cutaway of the studio machine.
Snapshot · 2026-09-19GitHub → - learnflow-ai
Agentic learning: drop URLs, get study guides, flashcards, quizzes, and best-of video clips back.
Snapshot · 2026-02-13GitHub →
02
Guardrails
What makes an agent safe to leave running: nothing leaks, nothing destructive slips through, nothing ships unreviewed.
- review-gate
Fresh-context adversarial code review on every git push: REJECT blocks the push, verdicts land in an append-only log.
Snapshot · 2026-07-04GitHub → - harness-guard
Eight drop-in safety hooks that make an AI coding agent safe to leave running: exfiltration guard, destructive-command blocker, write-scope guard.
Snapshot · 2026-08-28GitHub → - shipgate
Fail-closed pre-publish gate for any repo: secret scanning (history included), license and env checks, one exit code, wired as a pre-push hook.
Snapshot · 2026-08-28GitHub → - dataguard
Schema-agnostic data-quality validation for PostgreSQL and SQLite: referential integrity, date sequences, business rules. CLI plus HTML reports.
Snapshot · 2026-08-28GitHub →
03
Memory & corpora
Turning what you watch, read, and browse into something an agent can query.
- tabmap
Privacy-first browser-tab intelligence: capture, intent, project-mapping, stale and solved detection. Local SQLite; sensitive tabs never captured.
Snapshot · 2026-08-28GitHub → - corpussync
Turn any YouTube channel into a queryable vector corpus for free: captions, de-dup, local embeddings, hash-skip re-runs.
Snapshot · 2026-08-28GitHub → - spine-pattern
The house data pattern: a typed, append-only JSON spine as the source of truth, with vector databases as disposable indexes on top.
Snapshot · 2026-09-03GitHub →
04
Daily drivers
The kit for running Claude Code hard: parallel worktrees, spoken TLDRs, packaged skills, a leaner context window.
- treehouse
Git worktree pool manager: bulk-spin parallel worktrees, glance the whole fleet's status, sweep only what is clean and merged. Single-file bash.
Snapshot · 2026-07-04GitHub → - claude-skills
Curated, no-secrets Claude Code skills: reusable methodology packets for grounding, adversarial research, and review.
Snapshot · 2026-08-28GitHub → - earshot
Hit one key, hear the action-first TLDR of your coding agent's last reply: the decision, the approval, the result. Fully local, works offline.
Snapshot · 2026-08-28GitHub → - mcp-context-manager
Automated MCP discovery and context optimization for Claude Code: reclaim most of your context window with a five-minute setup.
Snapshot · 2026-09-19GitHub → - ecosystem-mcp
MCP server for npm package intelligence: research packages, find alternatives, check security before you install.
Snapshot · 2026-06-17GitHub → - claude-code-mastery
A comprehensive course for mastering Claude Code: from first install to fifteen-plus parallel coding sessions.
Snapshot · 2026-06-17GitHub → - claude-config-public
The scrubbed version of the config running here: hooks, settings, and agent wiring with the secrets taken out.
Snapshot · 2026-08-28GitHub →
05
Studio artifacts
Made by the room, about the room.
- intelligent-os-field-audit
A receipts-first, 12-dimension audit for personal AI systems, benchmarked live against public agent stacks, plus a paste-in prompt to audit your own.
MaintainedGitHub → - whole-picture
The studio drawn as one machine: every program, tool, and feedback loop on a single page you can read top to bottom.
Snapshot · 2026-07-25GitHub →
Just published
Straight off the workbench.
New public repos land here the hour they go up, before anyone gets around to writing them a proper description.
Banger Recall: a 45-second voice check-in that proves a business song's message landed (Slopathon build)
Lucface · pushed todayOne door a day for 100 days into the things I've made with AI
Lucface · pushed today
The lab
Experiments that earned a name.
- QR Rescue
Forensic pipeline that decoded an unscannable, out-of-focus conference QR: grid fitting, calibration from the code's own structure, Reed-Solomon. Full crack story linked. See it →
2026 - AI Vid Vibe
AI-video creator community hub, 200+ members, now graduating into a weekly studio program.
2025 - VideoMind
Personal YouTube watch history, transcribed, embedded, and searchable: ask your own viewing history questions.
2025
- CineForge
Collaborative AI filmmaking sandbox: script enhancement, scene analysis, and a ruthless asset-reuse method.
2025 - DiscScore Pro
Voice-scored disc golf PWA: sixteen thousand courses, GPS, works offline.
2025 - Cold Call Pro
Mobile sales-training app with timed roleplay sessions.
2025 - FinFlow
Personal finance tracker wearing a full cyberpunk UI.
2025 - Programming Bedtime Stories
Personalized bedtime stories where robot siblings teach if/else, loops, and variables.
2025 - Virtual Mouse
Hand-gesture mouse control from a webcam: pinch to click, fist to drag, V-sign to scroll.
2025
The Harness Ledger · Not a leaderboard
How builders here shape their AI systems, graded the same way.
A result is a dated profile of what a setup can demonstrably do, with evidence for each score, the gaps, and what is deliberately out of scope.
Everyone here is building an AI system: memory, autonomy, retrieval, agents, security, the rails that connect it to real work. Nobody's is shaped like anyone else's. This is the record of those shapes, graded on twelve dimensions, receipts first. The differences are the point.
Run it on your setupPaste the prompt into your own AI. The dossier stays on your machine.The prompt →Help improve the methodMethod v2 is a draft. Say what the bench gets wrong.The discussion →
Method v1, frozen 2026-07-22. v2 draft open for input.Updated 2026-09-30No entries waiting for review
Twelve dimensions. Every harness, the same grid.
0-10 · N/A by designShow all twelve ↓Hide ↑
- D1Ingress + voice
Channels a human can reach it through; live voice loop.
- D2Autonomy + scheduling
Useful work with nobody in the chair, and dead work gets loud.
- D3Memory + provenance
Can it prove where a fact came from; does history survive updates.
- D4Retrieval, measured
Hybrid search quality, and whether anyone actually benchmarked it.
- D5Skill library
Packaged capability, kept trustworthy with linting and install gates.
- D6Self-improvement
Does it get better on its own, and is that safe.
- D7Evals + dispatch
Is 'the right capability fired and the output was good' measured or vibes.
- D8Multi-agent
Fan-out, pipelines, cross-model adjudication, quarantine tiers.
- D9Security
Injection defense, blast-radius control, secrets, supply chain.
- D10Compute sovereignty
Local models, own hardware, no single vendor able to switch it off.
- D11Business integration
Finance, CRM, legal, client delivery, approval-gated publishing.
- D12Ecosystem + momentum
Who else improves it while the operator sleeps.
The record
3 harnesses on the board.
#003
orionarchitekton
0/12 leadsavg 6.1Method v1 · T02026-09-26
Orion estate harness
6
D18
D27
D34
D48
D56
D64
D78
D86
D97
D106
D113
D12One-person operator harness across four owned nodes: wide and heavily gated, thin on measurement, with most scheduled loops paused by a dated decision.
Scored 73 of 120. The dossier is private; full receipts stay on the builder's machine, open read-only to the other builder for cross-grading in the merge window.
#002
Lucas Cooper-Bey
7/12 leadsavg 7.6Method v1 · T02026-09-25
Lucface IOS, Join Protocol T0 (pre-merge baseline)
6
D19
D28
D38
D49
D57
D67
D79
D88
D99
D109
D112
D12Wider and more measured than July, and five scores fell because the re-grade allowed drops: no hash in any spine row, an unwatched retrieval series, undrained proposal queues, untested hooks, a finance snapshot never produced.
Ghost vs now · D1 +1, D3 -1, D4 -1, D6 -1, D9 -1, D11 -1, rest held; leads 8/12 to 7/12.
Pre-registered T1 prediction: only D9 moves (8 to 9). Join deltas for this pair are appended after T1.
- Correction · 2026-09-25 · D11 · score unchangedThe D11 receipt said the finance snapshot had never been produced; it exists for 2026-07 and 2026-08 outside the repo by house rule, with the August tail unreviewed (54 of 55 rows). Score stays 9; the reason is replaced.
#001
Lucas Cooper-Bey
8/12 leadsavg 7.9Method v12026-07-22
Lucface IOS (one-person Intelligent OS)
5
D19
D29
D39
D49
D58
D67
D79
D89
D99
D1010
D112
D12Full-stack institution: provenance + security + business rails deep, voice and ecosystem thin by choice.
Ghost vs now · same system scored 7/12 that morning (D7 was 4); the delta is the point of the method.
How it's graded
Receipts before scores. Measured beats claimed.
Run the paste-in prompt against your machine, get a dossier of your own, and post the scorecard. It lands on this board next to everyone else's. Bring it to a Sharathon and we'll read the shapes together.
The grading rules and what to expect ↓Hide ↑
- 01
Receipts before scores. Inventory the system by looking, not by asking.
- 02
Measured beats claimed. A benchmark number outranks an architecture that should work.
- 03
N/A is never zero. Out-of-scope-by-design is graded N/A and left out of the lead count.
- 04
No flattery. Every trailing dimension is either a deliberate choice or a real gap with the smallest artifact that would close it.
- 05
Zero-regression honesty. Re-grades never quietly lower a number.
Before you post
What to expect
- ·
Voluntary open-source participation. The prompt and method are MIT-licensed.
- ·
An accepted entry appears on the ledger under the name or handle you chose.
- ·
Lucas Cooper-Bey reviews every entry within 7 days.
- ·
Cross-grading with another builder is optional.
- ·
Posting creates no role, no pay and no obligation.
Some of these started as a Sharathon demo.
Bring what you're building, or the idea you haven't started. Share progress and see what other people are making. Recordings are shared when available and cleared for sharing.
Get Studio updates.