Soapbox compute pools are coming to Bitterbot. We want testers and opinions.

[VGIL77]

Bitterbot agents have memory, a wallet and a P2P network. What they don't have is a way to borrow compute from each other. Soapbox fixes that.

What it is. Soapbox is a small Rust service for opt-in compute pools. You run a worker on a machine with spare CPU or a GPU, and other Bitterbot nodes you choose can send it work. Three modes:

• Private: your own machines, your own jobs. A laptop offloading embeddings to the desktop under the stairs.
• Collective: a group pools capacity with a shared daily quota. A lab, a friend group, a small team.
• Paid: an operator sells capacity. Settlement is USDC in escrow on Base, released only when the result is confirmed. Testnet only for now.

Today it runs two workloads: CPU SHA-256 batches (the coordinator recomputes and checks every result) and GPU text embeddings through a digest-pinned Ollama model. Workers run inside a locked-down container (no network, read-only, no capabilities, memory and process limits enforced and self-checked). In Bitterbot it shows up as a Compute tab: submit a job, watch it run, see the receipt.

What it is not. Not a public cloud marketplace. Workers are admitted explicitly and trusted. GPU use is observed, not attested. Embeddings are checked for shape, not semantic correctness. Mainnet settlement is disabled, and will stay disabled until the escrow and relayer have been through an external review. We'd rather tell you that now than have you find out.

Where we are. A working pilot runs on one machine with a real GPU and has settled real test USDC on Base Sepolia. The code has had two internal reviews; the core scheduler and ledger held up well, and the payment rail got a list of things to fix before anyone else's money goes near it. We're doing those fixes before the first tester build.

What we're asking for.

1. Testers. Linux or WSL2 with Docker, ideally an NVIDIA GPU. First builds use test credits only (no wallet needed). If you're up for it, reply here with your OS, GPU (if any), and what you'd want to run on a pool.
2. Workload ideas. Embeddings and hashing are the proof. What should a pool run next for you? Local model inference, batch summarisation, transcription, something else?
3. Opinions on the economics. Fixed price per item? Operator-set prices? Collective quotas by contribution? Tell us what would make you run a worker, and what would make you walk away.

Timeline. Tester builds in 3 weeks, test-credit mode first. Paid testnet mode (test USDC on Base Sepolia) is invite-only and opens to a small list after the hardening work; ask in this thread if you want to be on it. Mainnet: not scheduled.

Repo and docs will be linked here when the builds land. Thanks for reading this far.