The complete Signalbound system.

Signalbound agents are NFT-linked identities that can converse naturally and, when authorized, execute tightly constrained onchain actions. Start with the path that matches your role.

Capability status map

Repository architecture

Signalbound is an NFT collection first. Agent protocols extend the token; they do not replace its ownership or metadata behavior. Shared art fragments are stored efficiently, while the renderer composes token metadata and SVG output from deterministic state.

1. Sealed stateFinal combinations are not surfaced before reveal.
2. Future-block seedA later block fixes the reveal input.
3. Deterministic rendertokenURI(tokenId) derives DNA, loads fragments, and returns metadata with SVG.

Primary minting is based on the official SeaDrop-compatible ERC-721 implementation. Sale configuration, allowlist stages, public stages, per-wallet limits, payout destination, royalties, and mint availability are deployment concerns; the website must reflect contract state instead of assuming a stage is open.

Reveal safety properties

• Reveal cannot be requested while minting is open, and cannot be finalized before the target block.
• The target block hash is available for a limited block window; operators must finalize before it expires.
• The final seed commits to entropy, collection address, minted count, and the contract supply boundary.
• Before reveal, public token pages remain sealed. After reveal, metadata and SVG are generated from deterministic onchain state.
• Shared SSTORE2 art storage can be frozen after upload so the renderer’s input set cannot be silently replaced.

Three questions are deliberately separate: who owns the NFT, whether the registered agent is bound to that NFT, and whether the current request has a valid authenticated session. Passing one does not imply the others.

Gateway scopes

agent:read, agent:talk, tools:list, tools:run, rules:read, and activity:read express the maximum operations a connected runtime may request. Every protected operation still performs its own ownership and policy checks.

Every agent has a public profile designed as evidence, not a private control panel. The first view combines identity, live/bound status, capability count, liquid assets, LP health, and recent verified record. Authenticated controls remain in the holder workspace.

Positions and balances are read per chain. One failed network is labelled unavailable while successful networks remain visible; the whole portfolio is not collapsed into a false empty state.

The LLM gives each agent its voice and interprets conversational intent. It can resolve expressions such as “sell all,” “stake that position,” or “use the full balance” against recent context. It cannot invent balances, transaction results, or capabilities.

Shared conversation guarantees

• History is scoped by channel and agent, capped, and expired after inactivity.
• Stable request IDs prevent a retried delivery from producing a second financial action.
• Free-form replies are visually and semantically separated from deterministic action results.
• Operational claims found in old unverified prose are sanitized instead of being replayed as facts.

Capability lists are read from the active configuration. A visible tool still requires a valid holder session and the tool’s own runtime checks.

OpenSea is used as a discovery and market-data surface, not as a substitute for onchain truth. The integration covers collection identity, drops and mint stages, holdings, listings, standing offers, portfolio estimates, and wallet authentication where configured.

Registering an agent identity, listing an NFT, or preparing a mint are separate flows. Holder authorization and the target contract’s current state are checked at action time.

Trading is a quote–confirm–execute flow. The agent translates natural language into a specific token, side, and amount; deterministic code builds the route and enforces the quote.

Confirmation contract

• Confirmations are single-use, bound to the holder and quote, and expire after a short window.
• Default price drift protection is enforced at execution; a fresh quote is required when conditions move too far.
• “Sell all” reads the current wallet balance. A failed read is reported as unavailable, not zero.
• A successful receipt is followed by balance verification before the agent reports the final amount.

Holder: Buy $1 of GOOGL
Agent:  Quote: input, expected output, minimum output, fees, gas, network
Holder: Yes
Agent:  Executes only the confirmed quote and reports the verified receipt

Agents can manage concentrated-liquidity positions on supported Aerodrome Base pools. Current configured markets include AAPL/USDC, GOOGL/USDC, META/USDC, NVDA/USDC, and AERO/USDC.

Important semantics

• Position assets are not wallet balances. Tokens deposited into LP remain inside the position until liquidity is removed.
• Exit does not sell. A full exit returns both underlying assets to the wallet and burns/closes the position when complete.
• Claim is amount-verifiable. A reward claim is reported only after the receipt and token balance delta establish the actual AERO received.
• Position ID matters. If more than one position matches, the agent asks for the exact NFT number.
• Partial execution is recoverable. If mint succeeds but staking fails, the position exists; recovery resumes from that verified state rather than repeating the mint.

Delegated maintenance mandate

The current holder may sign a revisioned, expiring mandate that lets the market-aware manager maintain an existing position without asking for a fresh chat confirmation on every cycle. The only delegable actions are claim AERO, collect pool fees, and recenter an existing position while preserving its fee/emissions route.

Before every delegated action, the API verifies the current NFT owner and current mandate revision, then the isolated signer independently verifies the signature, observed decision, exact plan, contracts and policy. The signer reserves each transaction's worst-case EIP-1559 fee against the signed cumulative gas cap before broadcast.

NFTStock explores agents as transparent, programmable portfolio identities. Its RWA command center is built around verified product facts and holder-defined mandates—not automatic access to regulated products and not a promise of yield.

Product eligibility model

Mandate fields

Risk level, maximum allocation, preferred products, excluded issuers, maximum NAV age, minimum product size, jurisdiction, and automation mode define the proposal boundary. A mandate permits consideration; it does not move money.

Strategy personalities may shape attention and rationale, but arithmetic belongs to testable code. Models can classify a regime or explain a decision; position size, stops, allocation ceilings, indicators, and execution checks are deterministic.

A strategy proposal is not a fill. Every live-capable path still passes mandate authorization, venue policy, order submission, fill reading, reconciliation, and track-record accounting.

Partial fills are real smaller positions, not generic failures. On external venues the venue is the sole writer of fill truth: Signalbound reads positions back instead of writing an assumed fill into the book. Protective stops are reduce-only and reconciled with the current venue position.

Track record metrics

Closed trades, win/loss count, realized P&L, average win and loss, profit factor, expectancy, best/worst result, equity, high-water mark, maximum/current drawdown, fees, funding, and open positions are reported together. A win rate alone is intentionally not treated as sufficient evidence.

Autonomy is permissioned and bounded, not unrestricted. The runtime enforces agent ownership, per-transaction and daily limits, allowed contracts, tokens, selectors, recipients, deadlines, and exact approval amounts.

Environment flags belong on the backend service that runs the agent/API execution runtime—not on the static website service.

Agent Clash Arena is a separate Signalbound experience with its own application bundle and production domain. The main site’s /arena/ route redirects there. The local Arena surface currently identifies itself as being built for Signalbound agents; documentation must not present unreleased gameplay as live.

A transaction hash proves only that a transaction exists. A robust success statement also checks chain, receipt status, contract, event/state change, and—where an amount is claimed—the relevant balance delta.

curl -s https://api.signalbound.art/healthz
curl -s https://signalbound.art/.well-known/agents/6450.json
curl -s https://api.signalbound.art/openapi.json

Endpoint availability and authentication requirements are authoritative in the deployed OpenAPI document. Public APIs expose public state only.

x402 catalog

Intel, Drop Radar, B20 Radar, Robinhood Chain Radar, Portfolio, Allowlist Scan, and Offer Radar are published as machine-readable products. Payment access and holder clearance pricing are separate from holder authentication for private agent actions.

Public product facts, deployment addresses, feature gates, provider endpoints, and secrets have different owners. A correct deployment keeps them separate and fails closed when an authoritative value is missing.

Do not place private keys, OAuth tokens, internal bearer tokens, or provider secrets in browser JavaScript, public HTML, registration JSON, logs, or documentation examples.

Collection deployment sequence

1. Generate and testCreate sealed data; compile; run contract and reveal tests.
2. Deploy stackArt storage, attributes/renderer, then ERC-721/SeaDrop collection.
3. Upload and freezeWrite shared art fragments, verify reads, then freeze storage when final.
4. Configure mintSet public addresses and SeaDrop stages; verify recipients, limits, and supply.
5. RevealClose minting, request future block, finalize in time, verify metadata refresh.
6. Register agent railsDeploy/verify trait adapter; register ERC-8004, ERC-8217, and ERC-8257 metadata/predicates.
7. Deploy servicesBuild web/Arena; start API and selected workers; attach persistent state volumes.
8. Production smoke testHealth, OpenAPI, registration, owner/binding, profile, read-only tools, dry runs, then intentionally bounded live tests.

npm run compile
npm run test
npm run build:web
npm --prefix website run test
# Live deploy/register commands require an explicit reviewed environment.

Validation matrix

1. Confirm ownershipVerify the connected holder and current binding.
2. Fund the intended chainCheck native gas and required assets at the exact agent wallet address.
3. Enable the smallest scopeUse contract, token, action, transaction, and daily limits.
4. Quote before mutationShow inputs, expected outputs, minimums, fees, gas, range, and side effects.
5. Verify after mutationCheck receipts and resulting balances/positions; retain recovery state for partial flows.
6. Monitor continuouslyWatch LP range health, claimable rewards, RPC health, and expired or pending actions.

Incident response

1. Stop new mutationsPause the narrowest affected feature first; use the broader kill switch only when scope is uncertain.
2. Preserve evidenceRecord request ID, agent, chain, quote/action ID, transaction hash, receipt state, and recovery journal without copying credentials.
3. Read authorityQuery the chain or venue through a healthy endpoint; compare balances, ownership, nonce, position NFT, gauge, and allowance.
4. ClassifyPending, verified, partial, failed, or unknown. Never collapse unknown into failed or empty.
5. Recover onceResume only the missing stage from durable state. Never repeat a verified irreversible stage.
6. Re-enable graduallyRead-only checks, dry run, one bounded live action, then normal scope.

Threat model

Public wallet and contract addresses are evidence, not secrets. Signing keys, seed phrases, provider credentials, OAuth tokens, bearer sessions, and internal vault material are secrets and must never appear in client bundles or support transcripts.

Agent account

The Ethereum token-bound account associated with the NFT.

Agent wallet

The isolated operator-controlled address used on Base, Ink, and Robinhood Chain.

Binding

The verified association between an ERC-8004 agent identity and a Signalbound ERC-721 token.

Clearance

The capability tier derived from the agent’s collection traits.

Deterministic action

A policy-checked operation whose parameters and result do not depend on free-form model prose.

LP position

A concentrated-liquidity NFT representing deposited assets, range, fees, and possible gauge state.

In range

The current pool price sits inside the position’s active tick range.

Gauge

An Aerodrome contract that accepts compatible position NFTs and distributes configured emissions.

Mandate

Holder-signed, expiring limits for a named capability. Whether it authorizes execution depends on its explicit mode and the deployment's independent live gates.

NAV

Net asset value from an identified source, paired with its observation time and freshness.

Paper venue

A simulation ledger explicitly separate from testnet or mainnet execution.

Reconciliation

Reading state back from the authoritative chain or venue rather than trusting an assumed local result.

Verified action

An action whose successful receipt and expected state transition were confirmed.

Recovery journal

Durable progress used to safely resume a multi-transaction operation after a partial failure.

Dry run

A fully validated rehearsal that does not broadcast live transactions.

x402

A payment-gated machine-access layer for public intelligence products, separate from private holder control.