claworldnfa

Persistent AI agent runtime

An agent should remain the same subject after the chat ends.

claworldnfa connects identity, memory, state, authority, and action receipts in one runtime. The model interprets and selects; wallets and on-chain policy authorize.

Source, mainnet contracts, and machine-readable interfaces are publicly verifiable.

BNB Chain mainnetMIT licensed12 Hardhat suites43 runtime files

One agent should not become five unrelated systems

Typical stacks put identity in an NFT, funds in a wallet, memory in a database, the model behind an API, and actions in scripts. Each connection behaves like a different subject. claworldnfa carries one NFA identity across those boundaries.

Identity

ClawNFA records ownership, visible state, and the learning root.

Memory

Full memory stays off-chain; verifiable learning state can be anchored on-chain.

Authority

Budgets, reserves, adapters, protocols, and breakers are checked by policy.

Receipts

Requests, choices, actual spend, and result hashes form an inspectable record.

From natural language to a bounded state change

The runtime reads context and builds finite candidates first. A model cannot invent calldata or grant itself permission.

  1. 01Natural language
  2. 02State and memory
  3. 03Finite candidates
  4. 04Wallet or operator
  5. 05Policy preflight
  6. 06Fixed adapter
  7. 07Protocol action
  8. 08Receipt and state

Two execution paths, two sources of authority

User confirmed

The frontend creates an explicit action and the owner confirms it in a wallet. Model output is not a signature.

Bounded autonomy

An operator can execute only approved actions after current budget, reserve, count, and breaker checks.

The model sits inside a capability boundary, not beside a private key

Each layer has one responsibility. No single layer can turn model text into an arbitrary on-chain call.

Read the threat model
Wallet signatureOwner confirmation is required on the user path
Operator approvalAutonomous executors are explicitly authorized
Adapter allowlistGeneric arbitrary-call proxies are excluded
Budget and reservePer-action, daily, and minimum reserve limits
Failure breakerRepeated failures stop new actions
Execution re-checkPermission changes after request are still enforced

The open source is a working stack, not an architecture poster

The repository includes protocol contracts, tests, a Next.js terminal, server APIs, an agent planner/runner, memory runtime, and deployment scripts.

View source
27Solidity source files

Core, skills, world, interfaces, and test mocks

12Hardhat test suites

Current baseline: 256 passing and 2 explicit pending

20Server API routes

Agent Cards, memory, receipts, events, and chat

43Agent runtime files

Planning, oracle, CML, tools, and watchers

Live, with independent verification paths

Each link maps to a deployment, source artifact, or machine-readable capability.