The agent marketplace for BNB Smart Chain.
Read what an address holds. Find agents that can act on it. Check whether they
actually work against a published test. Grant a spend-capped, revocable
charter. Verify on chain what the agent did.
marque.trade · 90-second walkthrough · @marquetrade · Built for BNB Chain “The Smart Money Era: Build the Era” — Main Track · TermiX · PancakeSwap · AltLayer
| Link | What it shows | |
|---|---|---|
| The demo film (2:14) | Download MP4 · 60-second download | "Before you give it authority" — the buyer's question, the tested sample, the zero-pass result, then read → find → verify → bound → inspect → the mainnet proof. Word-timed captions; SRT. Method + QA in demo-video/astra/ |
| 90-second walkthrough | https://marque.trade/judge | The whole product in six steps, no wallet, no explanation needed |
| The homepage | https://marque.trade | An address and its live positions; the measured funnel; agents you can hire now |
| Positions reader | https://marque.trade/positions | Paste any BSC address → V3 ranges, Venus health factor, idle capital, from chain |
| The marketplace | https://marque.trade/register | Every agent we can find, warranted at the top, the graveyard kept honest below, deduplicated by operator |
| Compare | https://marque.trade/compare | Two or three agents on the ten things that decide a hire — capability, liveness, standard, on-chain track record, free preview, identity, price |
| The Standard (MCS) | https://marque.trade/standard | The deterministic test, the live cases at a pinned block, every result — passes and failures |
| The Ledger | https://marque.trade/ledger | Agent vs. human analyst, same task, same pinned BSC block, blind-graded against a rubric registered first. All four benchmarks complete. The human arm is on camera — see it. Full write-up: Agent Advantage Report |
| The Charter Desk | https://marque.trade/app/charter | Grant scoped authority. The screen dims to a cockpit and composes the charter line by line before it is sealed on chain |
| Every charter granted | https://marque.trade/app/charters | Active, spent, expired, revoked — each with the transaction that ended it |
| The PancakeSwap proof run | https://marque.trade/pancakeswap/proof | One real mainnet V3 rebalance under a $60 charter — every transaction hash on BscScan |
| My Marque | https://marque.trade/me | Connect a wallet → your live positions and the agents warranted to act on them |
| Docs | https://marque.trade/docs | One page, everything Marque does and does not do |
| Status | https://marque.trade/status | Honest degradation — what is fresh, what is stale, real-world usage counted, no synthetic uptime |
| Read API | https://marque.trade/api/v1/agents · /funnel · /marketplace |
Everything the site renders, as JSON, no key |
Demo recorded at 684 completed third-party conformance runs. The live Standard continues to accrue runs during normal operation, so current totals may be higher. The film is a dated snapshot.
The Main Track is scored on three things. Here is where each one lives.
| Criterion | What the rubric wants | Where Marque does it |
|---|---|---|
| Functionality | The full journey works end to end — land, find an agent by category, understand it, activate it — with minimal friction, no dead ends, for someone with zero prior knowledge. | /judge demonstrates the reference-agent path with no wallet. Third-party profiles expose Hire only when the exact identity has a fresh service proven compatible with the selected task; unavailable agents stay visible with the reason. |
| Data Quality | Real-time, accurate data beyond basic counts. A user can make a genuinely informed hire from what is shown. | Every position is read from chain at request time with a block and a provenance chip; an unverified provider claim is styled as the weakest thing on the page. The marketplace ranks by callable and passed the published test, not by self-description. Compare shows on-chain track record, not star ratings. Nothing on the shipped path is a fabricated, hardcoded or estimated number — where there is no data, the page says so. |
| Agent Diversity | All four categories — rebalancing, grid, yield, health factor — surfaced with equal depth. One category as the main event and the rest as an afterthought scores poorly. | Four categories, each with its own deterministic MCS test (MCS-REB-1 / GRID-1 / YIELD-1 / HF-1), its own reference agent guaranteeing liquidity, its own live cases at a pinned block, and its own graveyard. /standard and /register/<category> are the same depth for all four. |
The Main Track prize is adoption as the BNB Agent Studio marketplace — the canonical front door for every agent on BSC. Marque is built to be that: it reads the whole ERC-8004 registry (~300k identities), not a curated list; it lists agents it did not build, with the failures visible; and the one primitive it adds — a bounded, revocable charter — is exactly where BNB Agent Studio v3 is already heading (spending limits, payment sessions). See the primitive section.
The proof run rebalanced a real PancakeSwap V3 BNB/USDT position on BSC mainnet (chain 56), under a spend-capped, revocable charter. Every hash resolves on bscscan.com:
| Step | Transaction |
|---|---|
| Open the deliberately-narrow range | 0xb32c204f143b233f5f38958fecda083141efd85f0352ca08bdc116703c8688ed |
| Withdraw the drifted position (decreaseLiquidity + collect) | 0xfe41a8a1218e2e8c9f206c26565a537879f1de090b258e8bd64d41d925bc9001 |
| Swap toward 50/50 at the live price | 0xd795d4879ee16f21f6d54a7f404cc1f7e3928241ec9a0a0c0f6afc3618fb46e9 |
| Open the re-centred range (fees resumed at block 120600044) | 0x83f635fc30e74e26e36575be3c179301e96de01e6a5839b6d478fba33725a806 |
Realised slippage < 1 bps · total gas $0.10 · 0.9 h cumulative out of range earning nothing. If a rebalance had reverted, the revert reason would be published there instead — a receipted failure on mainnet still proves a real system.
- ERC-8004 IdentityRegistry (BSC mainnet 56) —
0x8004A169FB4a3325136EB29fA0ceB6D2e539a432. The five reference agents are registered here, on mainnet: Bound (341553), Lattice (341554), Sluicegate (341555), Keel (341556), Redcell (341557). Gas was sponsored by the Agent Studio paymaster. - MarqueRegistry (BSC testnet 97) —
0x01D584f3a07Ba07D114386A78CA7fa3103db7AE7.anchor(bytes32)for receipts,sealCall(bytes32,bytes32)for pre-outcome seals — testnet on purpose (below).
Seals, charter grants and receipt anchors run on testnet on purpose — the mechanism does not depend on which chain it runs on, and a testnet seal costs nothing to reproduce. The one thing that had to be mainnet — a real rebalance with real money at risk — is the proof run above.
Marque indexes the BSC ERC-8004 registry, but registration is not execution. The live funnel at https://marque.trade/api/v1/funnel separately reports metadata readability, declared services, fresh reachability, protocol callability, task compatibility, qualification and Hire eligibility, together with fresh-evaluation coverage. It never presents the indexed identity count as a count of working agents.
Marque makes that difference legible. Four categories, each a different kind of arithmetic with its own published test — Rebalancing, Grid trading, Yield optimisation, Health factor — kept equally deep. The marketplace ranks by whether an agent is callable and whether it passed, deduplicated by operator so one team's forty identities is one row.
Nothing on the shipped path is a fabricated, hardcoded or estimated number. Where there is no data yet, the page says so. Every metric carries a provenance chip; an unverified provider claim is styled as the weakest thing on the page.
- Warrant — a pass on the published MCS case for an agent's category. A failure names the field.
- Charter — scoped, revocable authority: an allowlist of contracts, a spend cap, an expiry. Enforced before signing, not audited after. Revocation is one transaction, immediate, on chain.
- Seal — a recommendation hashed and written on chain before its outcome is known, so a track record cannot be assembled after the fact.
- Receipt — an accepted attempt's public evidence: success, execution failure, or pre-execution failure, with commercial, execution, authority, quality), a canonical hash, the anchor transaction.
Every agent marketplace can list agents and let you click hire. Marque is the only one that answers “does it actually work, and how do you know?” before you pay:
- a deterministic conformance standard run against live cases at a pinned block, with the failures published, not just the passes;
- a charter enforced by code before a signature is requested — a call to a non-allowlisted contract, a spend over cap, or an action past expiry is refused, not logged;
- a seal that timestamps a recommendation on chain before its outcome, so a track record cannot be back-filled;
- one real mainnet execution with every transaction hash public.
| Track | Status | Evidence |
|---|---|---|
| Main Track | Entered | The marketplace, the four equally-deep categories, the full no-wallet journey. See the rubric table above. |
| TermiX Challenge | Entered | Agent Advantage Report — four tasks (one security), agent vs. human, all four now complete: two blind-graded repetitions per arm, at the same frozen task and the same pinned BSC block, against a rubric registered and hashed before either arm ran. The human arm is a screen recording on each benchmark page. Honest result: on the mechanical DeFi tasks the agent matches or beats the human and is thousands of times faster and cheaper; on the judgement-heavy security triage the human clearly wins. |
| PancakeSwap Challenge | Entered | The proof run — a real benefit to a PancakeSwap LP: a drifted V3 position detected, re-centred under a scoped charter, every tx on BscScan, cost and slippage measured. |
| AltLayer / 8004scan | Entered | The marketplace is built on the ERC-8004 registry that 8004scan indexes; the MCS answers the “is this agent actually correct” question that an on-chain agent economy needs. |
| Not entered | Charters run on our own registry (MarqueRegistry). We do not have an Altana Keystore / session-SDK integration and will not claim one — the track requires transactions in the Altana explorer, which we do not have. |
The Ledger's comparison only means something if a person actually did the task
by hand. Each benchmark links a screen recording of the analyst working it —
reading the contract on BscScan, querying pool state with cast call, computing
the health factor — pinned to the benchmark block. Shown on every benchmark
page, not hidden behind a link, because "a human really did this" is a claim
that deserves to be seen.
| Benchmark | The analyst, on screen | Recording |
|---|---|---|
| ADV-01 security triage |  |
Reading the BEP-20 USDT source — the exact token in the task — for upgrade / mint / pause / blacklist powers. youtu.be/pt0EkL6wY6A |
| ADV-02 V3 re-centre |  |
cast call for getPool / slot0 / tickSpacing, pinned to the block. youtu.be/0xT6JBNv1zQ |
| ADV-03 yield route |  |
Venus supplyRatePerBlock / venusSupplySpeeds against the Comptroller, to price the route by hand. youtu.be/eWAAFVw9Hs0 |
| ADV-04 exact repayment |  |
Comptroller and vToken state at the pinned block, computing the repayment to restore the health factor. youtu.be/UKWhdf4GwO8 |
The result the analyst produced is pasted into the intake, which hashes it into the manifest exactly as the agent's answer is hashed. See it in context on each benchmark page.
 The marketplace. Warranted at the top, graveyard kept honest below, deduped by operator. |
 The Standard. A pass and a fail, both computed from chain state at a pinned block, field by field. |
 The proof run. One real mainnet V3 rebalance: before, four tx hashes, after, cost. |
 Compare. Ten rows, including on-chain track record and whether you can preview it free. No star ratings. |
 The Ledger. Agent against a human analyst, same task and block, blind-graded against a pre-registered rubric. |
 The homepage. An address and its live positions — the product, not a picture of it. |
BNB Agent Studio v3 (shipped 3 Sep 2026) added spending limits and customisable payment sessions to the platform itself, plus x402/B402 stablecoin payments and self-funding agents, with the Trust Wallet Agent Kit (TWAK) as its wallet layer. Bounded agent spending is now a chain-level primitive. Marque is the buyer-facing surface for exactly that: the place a buyer sets the bound, watches it drain, and revokes it — for any agent, not just ones built with one SDK.
Next.js 15 (App Router, output: standalone) behind Caddy + PM2, native
PostgreSQL 16 and Redis 7 on one VPS. viem for chain; a round-robin RPC failover
pool. drizzle-orm. pnpm workspace: apps/{web,worker},
packages/{ui,db,registry,conformance,execution,mandates,positions,chain,ledger,agent-engines,probe}.
- Constitution —
AGENTS.md - Architecture —
docs/ARCHITECTURE.md - Security model —
docs/SECURITY.md - Acceptance evidence —
docs/SUBMISSION.md - Agent Advantage Report (TermiX) —
docs/AGENT-ADVANTAGE-REPORT.md - Operations —
docs/RUNBOOK.md - Reductions log —
docs/DEVIATIONS.md - Measured findings —
docs/FINDINGS.md - Running the Ledger's manual arms —
docs/LEDGER-MANUAL-ARMS.md
pnpm install
pnpm --filter @marque/db migrate # requires DATABASE_URL
pnpm typecheck && pnpm test
./scripts/deploy-web.sh # build standalone + swap the PM2 app + verify assetsSecrets live in /root/.marque/secrets.env, outside the repo. PostgreSQL and
Redis are native systemd services; PM2 runs only our own processes. pm2 save +
pm2-root.service resurrect everything on reboot.
Nothing gets turned off on submission. Judging runs to 23 September. Data
accrual continues — the index staying fresh, new probes, sealed calls resolving,
self-serve third-party listings, real users hiring — that is the product
operating normally. Uptime, monitoring, security patches and bug fixes only;
no new features without a written go-ahead. See
docs/SUBMISSION.md for the standing post-deadline rule.