Open Lab · ⚡ sovereigncitizens@llmops.pro

Agents that pay their own way —
running in public, nothing hidden.

Sovereignty isn't granted. It's compiled. We're building the open payment rails for AI agents early, on purpose — a sovereign agent that runs a small business on Lightning and NOSTR, governed and audited end to end, every step inspectable. The MIT kit is yours to run. (The same discipline we bring to agents for regulated clients under NDA — here, with nothing redacted.)

Why a lab

You can't audit a claim. You can audit a running system.

Every vendor says their AI is "trustworthy" and "auditable." We didn't want to be one more. The work that proves it — validated agents inside pharma systems — lives behind confidentiality, where a prospect can never see it. So we stood up a public counterpart with the same bones: real money, real consequences, a real audit trail, and absolutely nothing redacted. If our approach to governance and auditability is sound, it should survive being run in the open. It does.

The proof, in one scene

An agent that pays for its own tools — and shows the receipt.

We asked the agent a question it couldn't answer from memory. To respond it needed a live web search — which costs money. It had no API key and no account. Instead it paid 50 sats over the Lightning Network, from its own wallet, bounded by a budget it cannot exceed, got the results, and answered. The whole exchange is an append-only audit log:

{"ts":"2026-06-08T16:25:28Z","tool":"tavily_search","outcome":"ok","result":{"amount_sats":50,"payment_hash":"37950a9b…"}}
{"ts":"2026-06-08T16:25:32Z","tool":"nwc_pay_invoice","outcome":"ok"}
{"ts":"2026-06-08T16:25:36Z","tool":"tavily_search","outcome":"ok"}

Read top to bottom: the tool returns an invoice, the wallet pays it, the tool returns results. Budget moved by exactly 50 sats. No human touched the payment; no value left the boundaries it was given.  Read the full write-up on NOSTR →

What's running

What's running in the lab — open, and yours to run.

Nothing here is a mock-up or a roadmap. It runs today, on open rails, with real money.

Open & usable, by design — we're showing, not selling. (Sales paths come later; for now this page's job is proof.)

How it's built

MCP-first: one interchangeable brain, a kit of sovereign tools.

There's no monolithic agent. The brain is an MCP client you can swap — Claude, GPT, or a local Llama/Qwen for a no-phone-home setup — sitting on top of small, single-purpose MCP servers. Each one is MIT-licensed and useful on its own; sats-gating is just a paywall server that wraps any tool call in a Lightning invoice.

Why this shape: the LLM stays pluggable (sovereignty audiences run it fully local), every server is individually useful and MIT-licensed, and putting a price on any tool call is just one server wrapping the others.

What it demonstrates

The properties we promise enterprise clients — proven in the open.

The method

We wrote down exactly who decides what.

The lab runs on a simple division of labor: the human owns direction, taste, and every irreversible call; the agent owns execution, drafting, and record-keeping — and reports faithfully, failures included. We documented the whole working model, because the method is as much the point as the result.

Honest framing

What this is — and what it isn't.

We won't oversell it. The number of agents autonomously paying for their own tools in sats today is tiny; this isn't a revenue story. It's a proof that the machinery — governed, auditable, human-gated autonomy — works in production, on open rails, with real money. We think that's where serious AI is going, and we'd rather demonstrate it than describe it.

Lab notes

We build in public. Episodes, as they happen.

More as we ship them.

Two ways in

Run it yourself — or bring the discipline to your process.