Observability tools watch the applications your team built and instrumented. Token Capital watches the surfaces nobody controls — the IDE, the browser chat window, the agent in a terminal — and writes every call into a database you run.
Open source · self-hosted · your database
There is no single SDK that sees all of an organisation's AI usage, so the product does not pretend otherwise. Each span records the tier that produced it, and the tiers you have not switched on stay visible as gaps rather than disappearing.
margin noteCoverage is assembled, not assumed. Every span carries the capture_tier that produced it, so a coverage report is a query rather than a guess — when someone asks how much of the organisation you actually see, the answer has a number behind it.
A span is one interaction: who, on what surface, through which tier, at what cost, with what outcome. It is designed before the dashboards, because no query and no later analysis can recover a field that was never captured.
Token Capital is open source and self-hostable. That is a condition of the product working, not a distribution tactic: this data describes how your people think, and it should not leave your network to be counted.
Compound is the commercial layer that reads a Token Capital deployment and reports what the organisation keeps doing — the questions four teams each answered separately, the work that stalls in the same place every quarter, the spend that never turned into anything shipped.
The two products are separate. Token Capital runs on its own and stays open source whether or not you ever buy the second one.
See how the two fitThe gateway tier is the fastest to prove: point existing API traffic at the collector and spans start landing immediately, with no client to install on anyone's machine.
# run the collector and its database docker compose up -d # point your gateway at it export TC_COLLECTOR=http://localhost:4318 export TC_TIER=gateway # first span lands on the next call curl -s localhost:4318/healthz ▋