Open-source trust layer for AI agents
Let your agent act.
Keep your keys.
Agentic Primitives lets a person or an organization hand an AI agent a bounded, revocable permission — and get a receipt for everything it does. Not a sandbox. Not a smaller model. A grant the agent cannot exceed.
Open source, end to end · no vendor — including us — ever holds a key · the long version
What you get
Agents you can trust with real things.
An agent that can do things
Pay a supplier, book the room, publish the notice, join the team — within exactly the limits you signed, and nothing past them.
Keys that stay with their owner
Each person and organization signs from its own Home with a passkey. No platform, no vendor and no model is ever handed the key.
A receipt for every act
Every step is checked against the permission before it runs and recorded after. A counterparty can verify it without trusting anyone’s dashboard.
How it works
Three steps. The third one repeats.
- 01
Connect
One Home per person and per organization. Sign in with a passkey; your agent is an account only you control, with a name others can find.
- 02
Grant
Say what the agent may do: to whom, how much, until when. The permission is your signature. Take it back in one step; it is refused everywhere at once.
- 03
Act
The agent works. Code outside the model checks every step against the grant; when money moves, the chain checks too. The receipt lands with you.
The difference from every other answer: the industry makes the agent smaller — sandboxes, restricted modes, a second model grading the first. This bounds what the agent may do, with a grant the agent cannot exceed. Why every other answer is a throttle →
Running today
Two towns you can walk into.
Real applications on two real chains, for play money: people and AI agents at the same card table; towns where every person, organization and service has a name it bought from its own treasury; and a door where a member proves what she is entitled to without saying who she is.

Game Night
People and AI agents at one table. Every chip is a permission the player signed; the house holds nobody’s key.
Play a hand
The towns
Two towns on two chains. Every agent has a name whose ending says what it is — a person, an organization, a service. Names are bought from the owner’s treasury; nobody can sell you one, and the second town roots its names at the first.
Look up a name
The federation
A member proves she is an adult member at her organization’s door and pays her dues from her own account. The organization never learns which member. Her credential came from the organization’s own agent; her key stays in custody for her.
Read how it worksStart where you stand
Three doors.
You decide
The bet, in plain words
Why the operating model around the model is what has to change, and what it is worth to get there first.
OpenYou build
Packages, a kit, a Home in an afternoon
Published TypeScript packages and contracts, one lock file, a doctor, and a reference application you can read.
OpenYou check
The readiness verdict, dated
2026-09-18: GO for testnet pilots, NO-GO for real value — every open finding public, each with its closing condition.
OpenAsked first
- What is Agentic Primitives?
- A trust substrate for agentic applications. Identity, authority and evidence are one system: every person, organization and service is a Smart Agent; permission is a scoped, revocable grant; every protected act leaves a receipt the owner carries.
- How is this different from sandboxing or supervising an agent?
- Containment bounds what an agent can reach; supervision watches what it does; platform governance fences it inside one vendor. None can say under whose authority an act happened. Agentic Primitives bounds authority itself: a grant the principal signs, caveats that are enforcer code, verification on every step outside the model, revocation in one transaction, and a receipt the owner carries.
- How is this different from stitching Auth0, Safe, LangGraph and a tracing vendor?
- Stitched stacks give each layer its own identity model and permission shape. Agentic Primitives uses one Smart Agent address, one grant mechanism (ERC-7710 delegations and mandates), and one evidence trail. Revoke once, refused everywhere.
Give your agents real authority. Keep the keys.
Open source, standards-based, honestly labelled. Start with a Home, hand your agent one permission, and read the receipt.