the-lyceum
v0.1.2
Published
The Lyceum — circuit breakers for AI. brake stops a dangerous action; redteam stops a one-sided conclusion; thrift stops the bill.
Readme
The Lyceum
Circuit breakers for AI. Three tools, done narrowly, that stop an agent before it costs you — one stops a dangerous action, one stops a one-sided conclusion, one stops you paying for the same file twice.
packages/
brake/ stops the action — CLI + MCP server + skill
redteam/ stops the argument — CLI + MCP server + skill
thrift/ stops the bill — CLI + MCP server + skill
server/ licensing, waitlist, admin console, marketing siteThe three tools
| | brake | redteam | thrift |
|---|---|---|---|
| Stops | A dangerous action | A one-sided conclusion | Paying for the same tokens twice |
| Catches | Data exfiltration, infrastructure attacks, credential access, destructive ops, money movement, impersonation | Overconfidence, unsupported claims, confirmation bias, false dichotomies, missing trade-offs, straw men, and three more | Re-reads of unchanged files, ANSI noise, repeated log lines, base64 blobs, lockfile hashes |
| Measured | ~1,370,000 scans/sec · p50 0.73µs | ~494,000 challenges/sec · p50 1.96µs | 78.5% saved on an agent loop, 97.6% of it lossless |
| Network on the hot path | none | none | none |
All three are always-on: the model calls them itself, with no slash command and no asking. All three are covered by one subscription.
The brake and redteam figures are measured on a single core after JIT warm-up,
and each package has a test/throughput.test.ts that fails the build if it
regresses — a performance claim nobody re-checks is a claim that quietly rots.
thrift's number needs its caveat stated with it. 78.5% is an agent
re-reading files while it iterates, which is where deduplication pays. On a
first pass over unique files thrift's lossless saving is zero — everything
it removes there is truncation you chose via --budget. Both numbers are in
packages/thrift/README.md, and thrift measure .
runs it on your own files rather than ours.
Running it locally
npm install
npm run build
npm testCloud benchmarks
.github/workflows/throughput.yml measures brake / redteam / thrift on
GitHub's cloud runners — ubuntu x64 and arm64 — nightly and on every
push to the three packages, and reports the actual measured calls/sec as a
step summary plus a benchmark-results.json artifact. Run the same
measurement anywhere yourself:
npm run build && npm run benchmarkThe script shares the corpora and best-of-N methodology of the CI throughput tests, and it also runs thrift's token-guard edge cases — the same JWT / image data-URI / response-JSON fixtures as the unit tests — so a cloud runner verifies behavior too, not just speed: a JWT whose claims got stripped would never show up in a calls-per-second number, so the script checks it and exits 1 if any guard is broken or any number drops below its floor. A "cloud number" is a measured fact, not a claim made on a laptop.
Then start the server:
LYCEUM_DEV_MODE=1 \
LYCEUM_ADMIN_KEYS="pick-a-long-random-string" \
npm run server| | | |---|---| | Site | http://localhost:3000/web/ | | Showroom (plans, setup guides) | http://localhost:3000/web/showroom | | Telemetry (live benchmark report) | http://localhost:3000/web/telemetry | | Admin console | http://localhost:3000/web/admin |
LYCEUM_DEV_MODE=1 bypasses payment so subscriptions activate without one.
Never set it in production — the server says so loudly at boot.
Configuration
| Variable | Purpose |
|---|---|
| LYCEUM_LAUNCH_MODE | waitlist (default) or open. Pre-launch, checkout is refused server-side and the site shows the waitlist. |
| LYCEUM_ADMIN_KEYS | Comma-separated admin license keys. Never read from the database — see LICENSING.md. |
| LYCEUM_JWT_SECRET | Signs sessions and license tokens. |
| LYCEUM_DB_PATH | SQLite file. Defaults to packages/server/data/lyceum.db. |
| LYCEUM_PUBLIC_URL | Public origin, used for checkout redirects and CORS. |
| LEMONSQUEEZY_API_KEY | Creating checkouts. |
| LEMONSQUEEZY_STORE_ID | Your store. |
| LEMONSQUEEZY_WEBHOOK_SECRET | HMAC secret. Without it every webhook is refused — which is the safe default, since that endpoint grants subscriptions. |
| LS_VARIANT_{SOLO,TEAM,SCALE}_{MONTHLY,ANNUAL} | Variant ids. Missing ones are a startup failure, not a 500 at checkout — a customer who clicks buy and gets an error is a lost sale you never hear about. |
| LS_VARIANT_WAITLIST_DEPOSIT | Variant for the refundable waitlist deposit. |
| LS_VARIANT_ADDON_CONNECTION | Variant for a single add-on connection ($100/mo). |
| THRIFT_BUDGET_TOKENS | thrift's default per-result token cap. Default 4000. |
How money and access work
- Someone applies to the waitlist — name, organisation, work email, phone. Consumer mail domains are refused with a route to take instead.
- They pay a refundable $50 deposit. It credits against their first invoice. It exists so the list reflects intent, not curiosity.
- An operator approves them by hand in the admin console. Nothing is automatic.
- On launch they check out through Lemon Squeezy, which issues a license key and emails it. Lemon Squeezy is the merchant of record, so VAT and sales tax are handled rather than left to you.
- The webhook mirrors that key locally, so every later check is a local lookup — an outage on their side must not stop a paying customer's agents.
Plans
| | Connections | Monthly | |---|---|---| | Solo | 5 | $99 | | Team | 10 | $299 | | Scale | 15 | $799 | | Enterprise | unlimited | by contact | | +1 connection | +1 | +$100 |
A connection is one install of one tool on one machine. Enterprise is deliberately not a checkout tier — past 15 machines you need a conversation about scale and procurement, and a number on a button there is either a lowball or a wrong guess.
Add-ons are priced above the tiers on purpose. Solo is $99 for five connections, so buying five add-ons costs five times what moving up a tier does. That is the intent: an add-on is for the team on Scale that needs a sixteenth machine, not a cheaper route into Team. They are stored separately from the plan, so changing plan never silently discards connections already paid for.
Licensing
Currently MIT. If you intend to sell this as closed software, read LICENSING.md first — it sets out what code protection can and cannot achieve, why moving the detection server-side would destroy the performance claim above, and what to do instead.
The short version: you cannot stop a determined developer copying code that runs on their machine. You can make copying not worth it, and that is a different and solvable problem.
