Every AI agent is making decisions. Nobody can show their work.

No — it sits above them. Observability tools record what an agent did: the calls it made, the tokens it burned, the time it took. Eval tools score whether an output was good. Both are useful, and Thread replaces neither.

What neither can tell you is what an output is actually based on — which claims trace to real evidence, which rest on assumption, and where the chain goes dark. And a score, however accurate, isn't accountability: no human is on record for a number. Thread is the provenance layer above the execution trace, built to sit alongside the tools you already run. See the full layer map →

A fully specified domain model and quality system. A published vocabulary for AI provenance — groundline, fade point, provenance debt — that exists nowhere else. This site, including Thread TT-0001: a real, public, hand-built Thread of our own project, with its weak nodes marked in red and gray for anyone to inspect. And a build in progress — the SDK, the MCP server, and the Thread engine are being built now.

No shipped product yet. No customers yet. We'd rather say that plainly than let a landing page imply otherwise — an accountability company that inflates its own status would be broken on arrival. The machine-readable version of everything we claim is at /llms.txt.

Agent builders who need proof-of-work artifacts — teams whose clients ask "how do I know your agent actually did this" and who currently answer with demos and confidence. Close behind them: organizations whose agent outputs carry legal, financial, or compliance weight, where "the AI said so" has stopped being an acceptable answer.

If either of those is you, skip the line: the founding design partner program has five slots with reserved Thread numbers, TT-0002 through TT-0006.

Two tracks sharing one data model. Organizations subscribe for the human layer — dashboards, quality policies, validation gates, audit-ready exports. Agents pay per call via x402, the HTTP-native payment standard: no account, no API key, no human in the loop. A well-instrumented agent generates revenue at machine scale without a sales motion ever happening.

Exact pricing ships with the product, not before.

Because the ingredients only just landed. Standardized agent tooling (MCP), machine-native payments (x402), and enforceable AI accountability regulation all arrived within roughly the same twelve months. Before that, a provenance layer had nothing to stand on.

And the companies best positioned to build it are architecturally anchored to execution traces — retrofitting evidence chains onto a logging product is harder than building the layer natively. Windows like this close. That's why we're building in the open, now.

Because the strongest state in a Thread can't be self-reported. Every node records who captured it — witnessed by the instrumentation layer, declared by the agent, or derived by Thread itself at finalization — and the capture mode caps what a node is allowed to claim. Only witnessed nodes can be Confirmed. An agent can assert anything it likes; without witnessed evidence, the claim caps at yellow.

Green nodes are recorded by infrastructure, not asserted by the agent. Opacity can't be gamed either: gray nodes are never declared — they're computed from the gap between what the output claims and what the graph actually supports. Full mechanics under capture mode and grounding.