Assay

Nansen tells you what smart money did. Assay tells you what that has been worth, and hands your agent today's signals with that measurement attached.

Assay sits between Nansen's data and an agent's order. It does four things:

1. Measured priors. Six Nansen flow segments at three holding horizons, measured on 12,781 real API calls with a method sealed before the first call. Every prior is in basis points, with a block-bootstrap standard error, an effective sample size, a p-value, and the multiplicity-adjusted bar it was judged against. No key needed to read them.

2. Live signals, gated. assay now pulls today's flows, turns them into signals by the study's own rules, and joins each one to its measured prior and a decision: trade (sized on the standard error), hold (real, but costs eat it), skip (no edge beyond drift), or refuse (never measured).

3. Live Hyperliquid positioning. assay positioning reads where every large position on a market liquidates and who holds it, and whether smart money sits on the other side of a crowd that is paying funding. A sampler builds the history on a clock.

4. Measure your own. assay study runs the same pre-registered instrument over your own tokens and dates, and assay score turns the ledger into verdicts.

It runs as a CLI, an MCP server for any agent (Claude, Cursor, anything that speaks MCP), and a Claude plugin. Zero dependencies: Node 18+ and nothing to install.

See it: gryps-finance.github.io/assay shows what the tool served on real runs (the day's 27 signals with their priors and decisions; the Hyperliquid positioning read for BTC, ETH and HYPE), the priors behind them, and how to run it. The study page renders all eighteen verdicts, the cost at which each would turn, what an agent that traded them would expect, and the window audit, straight from the engine's output. The instrument check runs the same page over a world with a planted edge, to show the instrument finds one when it is there.

Powered by Nansen API.

Quickstart (about five minutes)

git clone https://github.com/gryps-finance/assay.git
cd assay

# no key: what the study measured
node src/cli.mjs leaderboard
node src/cli.mjs prior "Smart Trader" 48
node src/cli.mjs audit

# with a key (app.nansen.ai/api), set in your shell only
export NANSEN_API_KEY=...            # PowerShell: $env:NANSEN_API_KEY = "..."
node src/cli.mjs now                 # about 100 credits: today's signals, each with its prior and a decision
node src/cli.mjs positioning         # about 65 credits: live Hyperliquid positioning
node src/cli.mjs demo                # both of the above in one paced take with the framing, about 45 seconds; what our recording shows

# the whole instrument against mock venues, no key
npm test                             # 206 checks

assay now prints one line per signal: the flow, the prior the study measured for that segment (here at 24 hours), the verdict, and the decision. This is a real run, on our host, 2026-09-27 17:35 UTC, 100 credits:

  ASSAY NOW  Nansen flows for 2026-09-26, 20 of 20 tokens read
  priors: daily arm of the study, scored 2026-09-20T00:26:42.834Z, 12 bps round trip, 18 hypotheses

  SEGMENT        TOKEN       FLOW  SIDE  24h PRIOR (net ± SE bps)   VERDICT       DECISION
  Fresh Wallets  WETH     +$8.33b  long  -19.6 ± 9.0  (n 5308)      fail-signal   skip
  Fresh Wallets  AERO    +$67.15m  long  -19.6 ± 9.0  (n 5308)      fail-signal   skip
  Fresh Wallets  AAVE    +$41.77m  long  -19.6 ± 9.0  (n 5308)      fail-signal   skip
  Fresh Wallets  WBTC    +$28.69m  long  -19.6 ± 9.0  (n 5308)      fail-signal   skip
  Fresh Wallets  ENA     +$18.03m  long  -19.6 ± 9.0  (n 5308)      fail-signal   skip
  Fresh Wallets  UNI     +$10.43m  long  -19.6 ± 9.0  (n 5308)      fail-signal   skip
  Top PnL        WBTC     -$7.44m  short -22.0 ± 15.1  (n 2420)     fail-signal   skip
  Exchange       WBTC     +$6.89m  long  -9.6 ± 9.7  (n 4912)       fail-signal   skip
  Fresh Wallets  LINK     +$4.99m  long  -19.6 ± 9.0  (n 5308)      fail-signal   skip
  Exchange       AERO     +$4.42m  long  -9.6 ± 9.7  (n 4912)       fail-signal   skip
  Exchange       ENA      -$2.86m  short -9.6 ± 9.7  (n 4912)       fail-signal   skip
  Exchange       UNI      -$2.83m  short -9.6 ± 9.7  (n 4912)       fail-signal   skip
  Fresh Wallets  CAKE     +$2.44m  long  -19.6 ± 9.0  (n 5308)      fail-signal   skip
  Exchange       PEPE     +$2.42m  long  -9.6 ± 9.7  (n 4912)       fail-signal   skip
  Exchange       WETH     +$2.39m  long  -9.6 ± 9.7  (n 4912)       fail-signal   skip
  Fresh Wallets  ONDO     +$1.26m  long  -19.6 ± 9.0  (n 5308)      fail-signal   skip
  Fresh Wallets  JUP      +$1.18m  long  -19.6 ± 9.0  (n 5308)      fail-signal   skip
  Exchange       LINK     -$1.15m  short -9.6 ± 9.7  (n 4912)       fail-signal   skip
  Top PnL        AERO      -$964k  short -22.0 ± 15.1  (n 2420)     fail-signal   skip
  Whale          ENA       +$818k  long  -48.1 ± 21.1  (n 1358)     fail-signal   skip
  ... 7 more, the largest flows shown first; every signal is in ledger/live/live-signals.json (--all prints them)

  27 live signals; 0 tradeable on their measured prior. Acting on all of them at 24h carries a measured net of -18.4 bps each at 12 bps round trip.
  Powered by Nansen API: tgm/historical-token-flow-summary, one call per token per day.

assay positioning reads the venue-wide screener and then the largest positions on the markets with the most open interest. Same host, same hour, 64 credits:

  ASSAY POSITIONING  Hyperliquid, sample 2026-09-27T17 (screener 2026-09-27T17)

  MARKET   FUNDING/yr        OI   SMART   CROWD  DIVERGE  LONGS LIQ <5%  SHORTS LIQ <5%
  HYPE          10.9%    $1.89b   +0.65   +0.04    +0.61           5.6%            0.0%
  ETH           10.9%    $2.95b   +0.80   +0.21    +0.59           5.1%            0.5%
  BTC           10.9%    $3.19b   +0.42   +0.15    +0.27          11.7%            5.0%

  ETH: Longs (liquidate below the mark): 3.1% of mapped long notional within 3%, 5.1% within 5%, 35.8% within 10%;
  the heaviest half-percent band sits at -9.0% to -8.5% holding 12.1% ($154.9m).
  ...
  Skew is by notional: +1 all long, -1 all short. Smart money's is the venue's full smart-money book; the crowd's is
  the largest positions pulled, excluding smart money. Measurements with an age, not a validated signal.

The full map also names who holds each band (smart money, whales); that part is served on your machine and not reproduced here, under Nansen's redistribution guide.

Related MCP server: Alpha Sentinel Market Intelligence x402 MCP Server

What the study found

Pre-registered on 2026-09-16, amended once on 2026-09-19 before any call, run on 2026-09-19: 20 tokens across Ethereum, Solana, Base and BNB (SOL and PYTH were excluded at resolution, and the reason is recorded), 18 months, one call per token per day. Six segments times three horizons is 18 hypotheses, so the bar is p < 0.05/18.

No segment's flows predict returns beyond drift at any horizon. All 18 pairs read fail-signal. The largest excesses:

The negative control behaves: Exchange flow is movement, not conviction, and it reads fail-signal at every horizon. The instrument is not inventing edge. On the synthetic world in test/, the same instrument recovers a planted 160 bps edge as 159.0 and a planted 81.3 bps edge as 82.7, so a real edge would have shown.

What this means for an agent: "smart money net-bought $2M of X yesterday" is not, by itself, a reason to size a position on X. An agent that trades those signals pays the round trip for nothing; an agent gated on Assay's priors doesn't take them. node src/cli.mjs leaderboard prints all 18 rows. FINDINGS.md has the rest.

A second finding, for anyone backtesting on Nansen's historical flows

A multi-day window from tgm/historical-token-flow-summary is not the sum of daily calls over the same days. The audit rebuilt every one of 1,558 weekly and 380 monthly windows from its days nine ways:

• for fresh_wallets the window is exactly its last day (1,957 of 1,957 pairs);

• for the other five segments no rebuild reproduces it; the closest is usually a short trailing window (two days at weekly resolution, seven to ten at monthly).

So: call at the resolution you will act on, and never sum short windows into long ones or read a long window as a sum. node src/cli.mjs audit prints it; AUDIT-CONVENTION.md has the method.

Use it from an agent

The MCP server speaks JSON-RPC over stdio and needs nothing installed. Readers need no key; the two tools that pull live (assay_now, hl_refresh) need NANSEN_API_KEY in the server's environment and have a hard per-call credit cap (500, set by ASSAY_MAX_CREDITS_PER_CALL).

Claude Code

claude mcp add assay -- node /absolute/path/to/assay/src/mcp-server.mjs

Claude Desktop, Cursor, or any MCP client (the client's MCP config file):

{
  "mcpServers": {
    "assay": {
      "command": "node",
      "args": ["/absolute/path/to/assay/src/mcp-server.mjs"],
      "env": { "NANSEN_API_KEY": "only if you want live pulls" }
    }
  }
}

As a Claude plugin: the repository root is the plugin (.claude-plugin/plugin.json, .mcp.json, and skills/assay/SKILL.md, which teaches the agent how to read the answers).

Ask your agent "smart money bought AAVE yesterday, should I go long?" and it gets the flow, the prior measured for that segment, and skip, with the reason. That refusal to act on an unmeasured signal is the design: an agent handed a raw signal treats confidence as edge.

More contexts (a Node library, a cron job on a server, reading the ledgers from another agent or a dashboard, running your own study) and the JSON shapes: docs/USAGE.md and docs/RUNNING.md.

How it measures

• A flow becomes a falsifiable claim. A net flow of at least $250k into a token (the segment's ten strongest that day) becomes "this token beats its own drift over the next 24, 48 or 168 hours", dated the first instant the day's data exists, with its entry price. No code path that forms a claim can see an outcome.

• Every claim gets a drift-matched control: the same token, direction and holding period, entered at a random offset within three days, from a fixed seed. Subtracting it removes the market and leaves the claim.

• Honest errors. Holding windows overlap, so sd/√n understates the error; Assay uses a block bootstrap and reports the effective sample size. Sizing uses quarter-Kelly on the outcome variance plus the estimate's variance, capped at 25% of the book.

• Two failures, never one. fail-signal (no content beyond drift) and fail-economic (real, but costs eat it) are opposite instructions. The second says "change the venue, horizon or expression", and it would be thrown away by any tool that just reports "no edge".

• Sealed in advance. PREREGISTRATION.md (sha256 bf3024fb33a53051…) and PREREGISTRATION-AMENDMENT-1.md (88328d592759d830…) were written before any call. The run's manifest (ledger/study/study-manifest.json, written as the run started) records both hashes, and the Nansen request id of every flow call on the ledger is in ledger/study/receipts.jsonl.

• Reproducible. Re-scoring the full ledger in a clean environment reproduces verdicts.json, study.json and sweep.json exactly, timestamps aside.

What ships in ledger/study, and what doesn't

The verdicts, the priors, the cost sweep, the window audit, the run manifest, the spend report, and a receipt for every flow call on the ledger (12,760, one per call key): its request id, the credits it cost, and a SHA-256 of the response. The raw responses stay with whoever made the calls: Nansen's redistribution rules keep raw API data with the caller, and the .gitignore does the same for your own runs. Nansen can confirm every call from its request id. Anyone with a key can re-run the study with assay study and compare.

Honest limits

• The priors are measured on 20 large tokens over 18 months at daily resolution. A token outside that universe gets its segment's prior with a note that it is an extrapolation.

• The Hyperliquid positioning read is a measurement with an age, not a validated signal. PREREGISTRATION-HL.md, sealed 2026-09-27 before the first scheduled sample, fixes six tests (crowding against funding, liquidation proximity against realised volatility, smart-money skew against returns) that score the sampled panel once 30 days of six-hourly samples exist, with a label-free control for each. The venue caps the all-traders positions pull at 1,000 a market, so the map is weighted on the largest positions covering about half of the book; every surface says so beside the number.

• The economics are a closed-form expectation, not a backtest. Slippage beyond the flat cost, capacity, funding and correlation between positions each make the real figure worse.

Built by

Gryps. Gryps builds institutional perpetuals on SEI; we built Assay to decide which on-chain signals are worth acting on, and publish it so any agent can make the same decision with the same evidence. Built for the Nansen Meridian Buildathon, September 2026.

MIT licensed. Data: Nansen API.