Sites
Compare
For internal testing. Every model answers the same prompts; you put their answers in order three at a time (or compare two), without knowing whose they are, and the votes rank them per use case. What you find stays here: people and the agent see each model's places on public boards (Models), not these.
Costs
What the platform spends: what each provider bills, from its own API, and where our own records say the money went.
Inference
What every turn did, by day and by what was in force when it ran: the model, the reasoning level, the prompts. The place to see whether a change did anything.
Jobs
Everything that happens without anyone in the room: the loops that do it, the scheduled refreshes they run, and the turns a full cell made wait. A loop that is stopped or throwing is the first thing to check — its work stops silently, and every other number on the overview keeps looking healthy.
Organizations
Cells
The droplets that run account containers, as their heartbeats describe them. A cell is bounded by RAM while sessions run: its run cap is (total RAM − 1.5 GB) / 600 MB, and a turn that finds its cell full waits in line rather than being refused.
Signup
Who may create an account, and who is let in. On the waitlist anyone may sign up, which shows how many want in, and you admit them a cohort at a time, sized to the fleet's room. An invite code lets someone in at once and works once.
Emails
Every email the platform sends, what sends it, and a test send of any of them. A test is the real template filled with sample values.
What each droplet's deploy watcher is doing: merge to main, CI goes green, and each droplet deploys the part that concerns it.
Talk, box by box: what each part runs on, where it runs, and whether it answered just now. The map is the call path; the table under it says what each technology is for.
Platform
Recent turns
| When | Account | Model | Tokens | Credits | Cost |
|---|
Accounts
| Name | Role | Organization | Verified | Credits | Workspace | Keys | Source | Created |
|---|
Personal details check
Workspace defaults
Platform-wide defaults for every workspace. They are kept by the platform's own workspace, which belongs to no account and wakes when these pages open; each save is applied to every active workspace, and a new workspace gets them when it is registered.
What is available to buy
Products, prices, coupons and promotion codes, held by Stripe. A price says what a pack costs and, in its metadata, how many credits it grants — the one thing this platform adds to Stripe's own model.
What it costs us
Packs the app is offering now
| Pack | Price | Credits |
|---|
Model costs (USD per million tokens)
A turn is priced from what the instance reports; when it reports no cost, the model's tokens are priced from these rates. Platform price is the provider rate times the markup; one credit is one cent.
Synced from the providers
| Model | Provider in | Provider out | Platform in | Platform out |
|---|
Fallback table
Matched by substring when a model was never synced; edited in control-plane/src/pricing.ts. Anything neither table knows gets the default rate.
| Model matches | Provider in | Provider out | Platform in | Platform out |
|---|
Shared learnings
What every workspace has taught the platform about tools and methods. Each workspace extracts lessons from its own finished runs and sends only the lesson — never the objective, transcript or files. Lessons are matched by meaning across workspaces; one reaches every workspace's prompts once enough different accounts have produced it independently, or when you promote it. Accounts that opted out in their privacy settings are never recorded.