Agents are about to act at a scale no human can review. The question stops being “is this text safe?” and becomes “who authorized this, and can you prove it?” Obol is our answer to the second question. Here's the path.
Obol v1 watched traffic and correlated it with kernel audit logs. Obol v2 gated tools behind approved plans and enforced them at the kernel. Both worked in the lab. Both taught us the same thing: detecting a bad action is harder than never granting the power for it.
A short, open spec for what an obol contains, how it's signed, how it attenuates, and how origin taint is expressed. Written to sit beside MCP authorization and OAuth token exchange, not replace them.
A small, auditable gate for MCP servers and coding-agent hooks, with a public test suite: every invariant on the How It Works page gets an attack that should fail. Released with the full lab protocol and raw data.
One small daemon on your machine that holds your real keys and mints coins for every agent you run — coding, email, home, finance. Approve from your phone. Read every receipt in one place.
When an agent pays a bill or signs up for a service, the receipt chain is the evidence of what a human actually authorized. The same record an auditor, a payment network, or an insurer can check without taking anyone's word for it.
Your assistant hands work to a travel agent, which calls a booking agent, which calls a payments agent. Every hop is a chance to add power that was never granted. Coins only narrow on the way down.
Scheduled agents act while you sleep, on instructions written days ago. Obol treats stored instructions as text, never as power — a saved job has to present a fresh coin to do anything real.
Payment protocols for agents are already shipping. A spend limit on a coin, bound to a single request, is the difference between a helpful purchase and a drained account.
Models can draft coins and explain refusals. They never decide one.
The gate runs where your agent runs. Your real credentials never leave your control.
We publish the limits alongside the features, and the attacks alongside the results.
The roadmap above is ordered, not scheduled. It moves when the work is solid.
If you build agents, run agents, or break them for a living, the draft is where your input counts most. Tell us what the coin needs to carry for your use.
SEND FEEDBACK →