signet

Signet is human-sealed autonomous agents for onchain finance. Your World ID is your signet, a proof of humanity stamped onto AI agents, giving them authority to coordinate and execute DeFi strategies through the Uniswap API. Without the seal, agents can't move funds.

The problem: autonomous DeFi agents are powerful but dangerous. Anyone can spin up a bot that drains wallets, front-runs trades, or manipulates pools. There's no way to know if an agent acting onchain is backed by a real, accountable human. The trust layer is missing.

Signet bridges this gap. A verified human seals their intent with World ID 4.0, and a coordinated team of AI agents handles everything else: market analysis, optimal routing, execution, and confirmation, all through the Uniswap Trading API. Users can even speak their trades. Voice input is transcribed and dispatched to the agent team in real time.

The key insight: proof-of-humanity isn't a login step. It's a permission gate on the agent's tools. The executor agent literally cannot call the swap function without a valid World ID session. This is identity-as-authorization at the agent level.

On verification, users claim an ENS subname (alice.signet.eth) and set configurable autonomy levels: supervised, semi-autonomous, or fully autonomous with hard spending ceilings. Agent metadata (risk profile, swap history, preferences) is stored in ENS text records, making agent identity portable and onchain. Agents can discover verified counter-parties by ENS name and coordinate direct trades, turning Signet from a personal tool into a coordination protocol.

Signet is built on a multi-agent architecture powered by Claude (claude-sonnet-4-6) using a think/finish tool loop pattern. A coordinator agent parses natural language intents from Telegram and spawns three specialized sub-agents, each with a fresh context and focused tool set:

The signal agent checks pool state, gas conditions, and recent price movement via the Uniswap API, then gives an opinionated recommendation (GO, WAIT, or CAUTION) rather than just reporting data. The planner agent queries the Uniswap Trading API for optimal routing and quotes, calculating slippage tolerance based on pool depth. The executor agent signs and submits transactions, but only if a valid World ID session exists. This permission gate is enforced at the tool level using World AgentKit.

World ID 4.0 proof validation occurs in the backend before agent tools are unlocked. On verification, users register an ENS subname under signet.eth with agent metadata stored as ENS text records (autonomy level, risk profile, swap count). The intent matching engine checks for verified counter-parties before routing through Uniswap pools, resolving matches by ENS name with reputation context.

A MiniKit 2.0 mini app provides an alternative entry point through World App, with a SignetRegistry contract deployed to World Chain that stores agent configurations onchain. The agent loop architecture is adapted from a production system (BrandMover) that has been running autonomously for months. The core engine.py pattern: Claude API call, tool dispatch, result injection, loop until end_turn. The Telegram bot interface, SQLite identity store, and sub-agent forking pattern were all ported from that foundation. Python is the primary language, using anthropic, python-telegram-bot, httpx, web3, and the ENS library.