apiary · a capacity pool

Your idle LLM hours are someone else's crunch.

Apiary pools members' agentic-coding capacity and self-hosted inference into one hive. Contribute what sits idle, earn a stake in the pool, and draw past your own ceiling when it's your turn to burst. Routing, metering, and failover are handled by an engine that already runs in production for its founding operator.

Join the waggle list → How it works ~2 minutes · no wallet needed

the loop

Contribute. Earn. Draw.

Individual capacity is bursty — subscriptions idle all week then hit their ceiling mid-crunch; GPUs hum at 4% duty cycle. Pooled, the peaks cancel out. Tokens keep score of who smoothed whom.

Contribute

Bring a metered coding-agent plan or a box that serves open models. The pool's brain meters delivered capacity — the router and the meter are the same process, so there's no self-reported number to game.

Earn

Contribution mints your stake — earned on delivery, never on pledge. Governance of the pool (which sources join, how lanes are priced) belongs to the members who feed it.

Draw

Crunch week? Your tooling transparently rides pooled capacity, per model lane — heavy lanes burst onto the pool while light lanes stay home. If the pool vanishes, everything degrades to exactly your own setup. It can never break you.

The engine is real — extracted from a production single-operator deployment

lane=fable   INJECT→9f3ac1d2 [pool-b]     ← heavy lane riding pooled capacity
lane=haiku   passthrough [home]           ← light lane stays on home base
lane=local   qwen3.6:35b → 127.0.0.1      ← self-hosted model, same picker
🔀 SWITCHOVER[fable] reason=scoped-cap-83% tier=routine (headroom at destination)
429 → transparent retry on alternate source · client saw one 200

waggle list

Tell us what you'd bring to the hive.

We're forming the founding cohort and writing user stories from real answers, not guesses. Rough numbers are fine. Only the email is required.

You

What would you contribute?

Pick any. "Nothing yet" is a fine answer — drawers make the pool useful.

a coding-agent plan self-hosted GPU inference dev time on the engine nothing yet — I'd draw

What would you draw?

burst coding capacity in a crunch cheap bulk inference model variety without more subscriptions just curious

Calibration

The goldmine

What's the most painful thing about your current LLM setup? Be specific — this is what we build from.

You're on the list. 🐝

We'll write to you when the founding cohort forms. Until then, the engine keeps humming — and your answers just became a user story.