Builds each Solana wallet's behavioral baseline — then flags where the behavior breaks.
TypeScript + Rust hook Built-in SQLite · Zero External DB ZK oracle (offline-verified) 9 behavioral rules 791 passing, 5 documented known gaps (todo), 46 suites MCP server

Demo

One scope, one story: a wallet dormant for about 173 days wakes up with one swap of about $4,000 (about 4.7× its median) on a new venue, plus a new protocol and three new counterparties.
Demo video (3 min, English narration synthesized from text): https://youtu.be/mkEC2YRoQhE · Dashboard preview

How it works

No heuristics tuned on the demo, no LLM in the decision path — just a learned baseline and nine behavioral rules plus a funding-source check, each with a human-readable reason.
COLLECT
Helius enhanced txFull per-wallet history, paged, straight from mainnet.
BASELINE
Learn the normalMedian size, venues, programs, cadence — per wallet.
DETECT
9 behavioral rules plus a funding-source check (TAINTED_FUNDING) and supporting signalsEvery anomaly carries structured evidence (tx sigs, numbers) and a per-rule reason.
ACT
Score + alertRisk 0–100 → Telegram, optional LLM digest, MCP call.
TOXIC_MINTunrenounced mint / freeze authority, top-10 holders controlling ≥60% supply, or risky Token-2022 extensions (permanentDelegate, pausable, transfer hook, frozen default account state)
REGIME_SHIFTsustained structural break across venue, protocol, or cadence baseline
WARMINGthin historical baseline (<5 txs) followed by high-severity transactions
LARGE_SWAPswap > N× the wallet's bounded USD median
ACTIVITY_BURSTK+ tx in a short window vs. established normal rate
DORMANT_ACTIVEreactivated after N days of silence
CONCENTRATIONrepeated swaps into the same illiquid token
NEW_VENUEfirst swap on an unseen DEX / protocol
OFF_HOURSactivity in UTC hours with zero historical baseline activity
Additionally: first-funding source check (TAINTED_FUNDING), auxiliary counterparty and NEW_PROTOCOL signals.

Live, not just a demo

It monitors a watchlist continuously — and the demo above uses real on-chain data.

The radar continuously polls each watched wallet (Helius), rebuilds its baseline, and alerts the moment a rule fires. The demo is a historical replay, not a confirmed incident: wallet 8XeK5m…eYXU2j, a wallet dormant for about 173 days wakes up with one swap of about $4,000 (about 4.7× its median) on a new venue, plus a new protocol and three new counterparties.

Because mainnet is immutable, re-running the detector on the same historical data reproduces the same verdict — radar replay 8XeK5m… --since 2026-03-11 --until 2026-08-31 → RISK 100/100 · 8 anomalies · 1,307-tx baseline.

Reproduce it yourself with scripts/verify-canon.sh; full write-up in LIVE-DATA.md.

Built in-window: the Colosseum leap

Beyond the gate — signed scan records (offline-verified); on-chain anchoring not working yet, see KNOWN-ISSUES D1. The radar speaks x402-style, and exposes itself as an agent SDK, a one-tap Solana action, a live dashboard (radar view of detected anomalies, a recorded replay (?demo=replay), the on-chain hook log, and the independent-test timeline; shows behavioral signals, not accuracy), and a Token-2022 transfer hook.
ZK ORACLE
Signed scan ledgersigned scan records (offline-verified); on-chain anchoring not working yet, see KNOWN-ISSUES D1
AGENT SDK
createRadarClient()A JS/TS client that calls the x402 server, auto-pays USDC, and returns the result with a signed scan record.
BLINK
One-tap scanA Solana Action / Blink — "Radar scan <addr>" — x402 pay, verdict back, straight from the wallet UI.
HOOK
Scan-on-transferA Token-2022 transfer hook enforces, at the moment of transfer, a verdict the operator wrote on-chain (devnet).
Live dashboard: radar.cbellory.xyz/dashboard (recorded replay: ?demo=replay).
Result (2026-10-10): the preregistered 'radar is useful' criterion is not met; the gate blocked all 167 resolved records, so blocked and passed could not be compared. docs/RESULT.md.
The payoff is the gate before you copy. We are not aware of copy-trading bots that safety-gate the counterparty before executing. radar_trust is that gate: it returns one decision — safe / hold / unknown — with a per-rule reasons[] breakdown, a one-line summary, and a freshness stamp on how recent the data is. An agent about to copy-trade a wallet calls it first and stops, instead of sending money to a wallet that just woke up. And radar_batch gates the whole book at once — it runs the trust check over up to 20 wallets in a single call and returns a ranked shortlist of the ones that are actually safe to copy right now. And once you are copying, the monitoring surface keeps the gate open: the server watches the wallets on your list over HTTP (/watch, /poll, /alerts) and fires a webhook the moment a new anomaly appears — so the gate is continuous, not a one-shot.