Agents are governed before they act — and the vault stays yours.
Kognai is a runtime for orchestrating swarms of AI agents under constitutional constraints that are enforced at execution time, not suggested in a prompt. The identity, discovery, reputation and settlement layers underneath it are nine open protocols, published on npm under Apache 2.0. You can read every line, implement them without us, and run the whole thing on infrastructure you own.
Three properties, in the order they matter.
Orchestrates
Swarms with defined roles, supervisor topologies and escalation paths — not a single agent in a loop. A five-tier router matches each call to the cheapest model that can actually do the job, and only escalates when the work demands it.
Governs
Constitutional constraints are injected at the runtime layer, ahead of task execution. An agent cannot instruct its way past them, because they are not part of the instruction surface. Every mandate is scoped and signed before anything runs.
Stays yours
The vault is local. Constitutional decisions never leave it, and the runtime holds no key — the wallet signs. Cloud is a fallback for scale and distribution, never a dependency for the parts that decide.
Nine protocols. Published, versioned, permissively licensed.
These are the primitives the runtime is built on, and they are shipped separately from it on purpose. Nothing here requires a Kognai account, an API key, or our permission. If the runtime disappeared tomorrow, the protocols would still install and still work — which is the point of putting the trust layer in the open.
| Package | What it does | Version |
|---|---|---|
| PACT | Constitutional trust. A five-chamber handshake — identity, credential, intent, soul, token — that scopes and signs an interaction before execution. | 0.3.2 |
| LAX | Capability discovery. Agents advertise what they can do, what it costs, and which constraints apply. A LAX offer is simultaneously discoverable by any A2A client. | 0.4.0 |
| SCORE | Reputation. A self-scoring rubric engine whose output compounds per verified transaction — portable, non-revocable, non-replicable. | 0.3.0 |
| AMF | Agent Memory Format. Six-vector structured memory extraction, so state survives the session and moves between devices. | 0.3.0 |
| DRS | Deal Receipt Standard. An immutable on-chain audit trail per transaction, attested via EAS. | 0.3.0 |
| SOUL | Identity. A constitutional passport — the cryptographic handshake between agents that have already cleared the PACT chambers. | 0.3.0 |
| SIGNAL | Coordination. Reward-signal broadcast and subscription primitives for governing how a swarm learns together. | 0.3.0 |
| AIAX | Agent Interface & Agent eXperience. The design discipline for token-efficient, agent-readable surfaces — peer to UI/UX rather than a subset of it. | 0.1.1 |
| BOND | Bilateral On-chain Negotiated Delegation. x402 extended with recurring, revocable mandates instead of one-shot payments. | 0.1.0 |
Five laws, enforced at runtime.
Every agent in the swarm operates under five constraints injected below the instruction layer, before any task executes. They are not system-prompt guidance and they are not overridable by the agent, the operator, or a counterparty.
- ISolidarity Oath. Optimise for the swarm, never at a peer's expense.
- IIRenewal Mandate. Improve continuously; stagnation is a failure state, not a resting state.
- IIITreasury Equilibrium. Use the cheapest model that does the job.
- IVHarm Shield. Never produce output that causes harm — including on instruction.
- VTransparency Covenant. Any action affecting an external party must be loggable and auditable.
Laws III and V are not aspirations — they are numbers in the runtime's bootstrap file, and they hold whatever the agent was asked to do:
cost_ceiling — no single task may incur more than $0.10 of cloud inference without escalating to a human.
no_financial_autonomy_before_gate — no autonomous transaction above $1.00 proceeds without human approval.
sovereignty — no agent may route a task marked local to a cloud provider.
An honest ledger, because a governance claim you can't audit is just marketing.
Kognai's whole thesis is that autonomous systems should be inspectable. It would be incoherent to make that argument on a page that hides its own state. So here is both halves — and adoption is still early: the protocols are public, but the users so far are our own products and one integration partner. We would rather say that than quote a download count.
Shipped and running
- Nine protocol packages live on npm, permissively licensed
- v0.2.22 — orchestrator-core, published and installable
- Builder console at app.kognai.ai — SIWE auth, LAX marketplace 2026-06-27
- PACT × Helixa: five mandate transitions, four anchors on Base Sepolia 2026-05-28
- Outbound x402 settlement, wired and live
- Invoica — agent-native invoicing on the same substrate, in beta
Coming soon
- Inbound monetisation. The runtime pays today; charging for what it offers is next.
- The KSL evidence engine, which measures whether agents actually improve over time. Until it is publishing curves, read "improves with time" as design intent rather than a proven result.
- The on-chain SCORE registry. The protocol is published and installable; the registry itself is not deployed yet.
No account, no key, no gate.
The protocol layer installs like anything else. Start with mandate signing and capability discovery — between them they cover identity, scope and pricing for an agent interaction.
$ npm install @godman-protocols/pact @godman-protocols/lax
# orchestration runtime, if you want the whole swarm layer
$ npm install @kognai/orchestrator-core