Teach any AI agent to read live crypto whale flow like a trader. Six skills, no API key, auditable curl calls only.
Agents are good at reasoning and bad at knowing what whales did four seconds ago. These skills fix the second half: real executed whale trades across 15 exchanges (Coinbase, Binance, Kraken, Bybit, OKX and more) plus Ethereum, Base and Arbitrum DEX swaps, a Smart Money Radar for unusual flow, real liquidation trades, and signal outcomes, all through the free keyless CoinLobster public API.
They also carry the part data alone can't: methodology. How to read a cluster of buys vs one block trade, when counter-trend whale buying matters, how CEX/DEX divergence works, what a liquidation cascade configuration looks like, and how to check a signal's actual recorded track record before trusting it.
Ask a stock question and an agent can lean on decades of tooling. Ask "what are whales doing with SOL right now" and most agents either guess, hallucinate a number, or quote a headline about something from last week. The underlying data (executed trades, size-filtered, across venues, in one stream) exists at coinlobster.com, free, with no login. What was missing was the packaging: a format agents natively understand, with the reading discipline built in.
Design rules these skills never break:
- Read-only. Every call is a plain HTTP GET you can read in the SKILL.md itself. No bundled binaries, no opaque scripts, no writes.
- Descriptive, never predictive. The skills report what whales did and what price did after past signals. They are written to refuse forecast language.
- Evidence attached. Track-record numbers come with their sample size and recorded window ("9 of 14 recorded days"), never as bare percentages.
- Honest when dark. If a feed is degraded, the skills say so instead of interpolating.
Claude Code (as a plugin). The bundled .mcp.json wires the hosted MCP server at the same time.
/plugin marketplace add coinlobster/agent-skills
/plugin install coinlobster@coinlobster
Any skills-aware agent (Claude Code, Cursor, Codex, Copilot, Windsurf, Cline, Goose, Gemini CLI, 70+ others):
npx skills add coinlobster/agent-skillsThe data, with zero install. Point any MCP client at the hosted server, or just curl the keyless endpoints:
claude mcp add --transport http coinlobster https://coinlobster.com/mcp
curl -s "https://coinlobster.com/api/public/whale-radar?window=4h"By copy. Each skill is a folder with one SKILL.md and no dependencies. Copy the folders you want out of skills/ into your agent's skills directory (.claude/skills/, .cursor/skills/, or wherever your runtime looks), or paste a SKILL.md straight into context. That is the whole format.
claude.ai (web/desktop): zip a skill folder, upload at Settings > Capabilities > Skills. Needs network access enabled. (Not available in the Claude API code-execution sandbox: no network there.)
Most people wire these into a trading setup next to their own signals. The pack is organized around the six questions traders actually ask an agent, each answerable here keyless and (checkably) nowhere else end to end: the forced-vs-organic autopsy after a move, the crowding-and-carry pre-entry check against the coin's own recorded days, whale-direction-vs-my-side, where liquidations actually fired (trades, not a model), whether a flag has historically been worth anything (scored, never backfilled), and standing watches defined on the coin's own percentiles instead of static thresholds. Everything stays descriptive: the skills report state and recorded history with sample sizes; decisions stay yours.
The core skill: what whales are doing on a coin right now. Live executed trades (side, size, venue, on-chain vs CEX), hourly net-flow history, and the reading discipline: clusters beat single trades, counter-trend flow beats with-trend flow, and a $4M buy into a falling market is a different fact than a $4M buy into a rally.
"What are whales doing with BTC right now?" · "Is anyone accumulating SOL?" · "Big ETH sells in the last hour?"
Under the hood: GET /api/public/crypto-whales, GET /api/public/onchain-whales, and GET /api/public/whale-flow-history for the hourly buy/sell series on the same grid as hourly price. Free floors: every $1M+ trade on any coin in real time, $250K+ DEX swaps.
Which coins have whale volume genuinely out of line with their own baseline (not a global threshold: $3M on a small cap means more than $30M on BTC), and what actually happened after past flags. The skill insists on the second step: before treating a flag as meaningful, pull its recorded outcomes and report the sample size.
"Which coins have unusual whale activity today?" · "Has this signal been worth anything historically?"
Under the hood: GET /api/public/whale-radar, GET /api/public/receipts/{type}/{coin}, flag outcome history.
Real liquidation trades (not estimates): who is being wiped, on which side, at what price, across venues. Cascade detection (liquidations chaining across exchanges) and squeeze configurations (crowded side + forced flow against it).
"Are longs or shorts getting wiped on BTC?" · "Any liquidation cascade risk right now?"
Under the hood: the liquidations MCP tool (keyless, any pair) and GET /api/public/liquidation-history?pair=....
A fixed diagnostic order so the agent doesn't cherry-pick: magnitude first (is the move even unusual), then forced flow (liquidations), then whale flow, then positioning (funding/OI vs the coin's own recorded history), then, only last, narrative.
It also carries the forced-or-organic autopsy for post-move questions: liquidation events sequenced against price (did forced flow lead the move or follow it), OI purge vs build, funding flip, and whether size spot-bought the low. The three series only this API holds together keyless.
"Why is SOL moving today?" · "Give me a BTC market snapshot." · "Why did BTC just dump 5%?" · "Was that flush forced? Did whales buy it?"
Under the hood: market snapshot + baseline endpoints, including percentile judgments against months of recorded daily history ("funding in its top 5% of 127 recorded days").
The pre-entry and in-position picture: crowding as percentiles of the coin's own recorded days, carry cost in dollars, whale direction vs your side, forced-flow map, watch-set conditions, percentile alert recipes.
"I'm about to short ETH, what does the data show?" · "I'm long BTC and funding spiked, what should I watch?"
Not analysis: building. Teaches an agent the full REST surface and the MCP server so it can wire whale alerts into a Discord bot, log flow to a database, or build a dashboard. Covers auth tiers, rate limits, credit costs, and the exact response shapes.
"Add live whale alerts to my Discord bot" · "Build a script that logs BTC whale net flow hourly"
Wallet identity is Nansen/Arkham territory, live Hyperliquid positions belong to the HL trackers, and decade-deep on-chain metrics belong to Glassnode: use them alongside this. What is checkably unique here, keyless: executed cross-venue whale trades (not transfers), real multi-venue liquidation trades (a famous competing heatmap is, per its own docs, a model), a signal ledger scored live and never backfilled, and percentile judgment against each pair's own recorded days.
| Exchanges | 15 CEX, executed trades only |
| Chains | Ethereum, Base, Arbitrum (DEX swaps) |
| Whale floors | BTC $500K+ spot / $1M+ futures, ETH $1M+ / $2M+, top majors $400K+ / $800K+, everything else $100K+ / $200K+. DEX swaps recorded from $100K on Ethereum, $25K on Base and Arbitrum; the keyless views show $250K+ |
| Extras | Smart Money Radar, real liquidations, funding/OI vs recorded history, signal outcome log |
| History | Daily recorded series going back months, with percentile baselines and next-day outcome stats |
Everything above works keyless. Free real-time: BTC/USD plus every $1M+ trade on any coin; other pairs on a 15-minute delay.
A free key is a one-off 1,000-credit grant (no card, coinlobster.com/developers) that unlocks the metered REST endpoints at /api/ai/v1/*, roughly 20 to 100 calls in total. It does not lift the MCP caps. Full MCP depth takes an active developer plan (Builder $29/mo = 100k credits, Growth $99 = 500k, Scale $249 = 2M). One plan covers both: full-depth MCP and the monthly REST credits, on one shared balance. A REST call costs 10 to 50 credits; an MCP tool call costs 2x its mirrored REST endpoint, 40 to 100 by tool. Keyless calls, discovery, failed calls and no-data answers cost nothing.
CoinLobster Pro ($9/mo) is a separate product: the coinlobster.com website (live feed, alerts, whale radar). It does not include MCP depth or API access.
Both, ideally. The MCP server (hosted at https://coinlobster.com/mcp, keyless) gives an agent structured tools; these skills give it the methodology, and they work in agents without MCP support via plain HTTP. When both are present, the skills prefer the MCP tools and fall back to REST.
Tool count, honestly: tools/list currently returns 25 read-only tools (24 canonical plus whale_tape, a deprecated alias of whale_trades). Call tools/list before relying on any tool: the set grows, and this README is not the source of truth.
Keyless is shaped, not blocked. Every tool answers keyless, but several answer less: market_snapshot is BTC/USD only, whale_radar withholds magnitudes on every coin outside the free pair (BTC), whale_tape is 15-minute delayed off BTC/USD (except any $1M+ trade, which is always live), market_history caps at 30 days, whale_context withholds USD magnitudes, liq_zones covers free pairs only, and squeeze_score needs depth outright.
Is this really free? Yes. The free tier is the distribution layer of a paid product (developer plans, from $29/mo); the skills only use public endpoints.
Will it give trading advice? No. The skills describe flow and report recorded history with sample sizes. Nothing in them forecasts.
Why do numbers say "recorded days"? Because that's the honest window: stats are computed over the days CoinLobster actually logged, and the count is always shown.
Rate limits? Keyless REST (/api/public/*) allows 60 requests/min per IP; the 61st returns 429. The keyless MCP is rate-limited too, on a layered per-IP and per-tool budget that is not published as a fixed number: honour the Retry-After header on a 429 rather than hard-coding a rate. Both are fine for interactive agent use. The free key is a one-off 1,000-credit grant (about 20 to 100 REST calls in total), enough to try the keyed endpoints, not enough to run a polling loop. Sustained loops and bots need a developer plan (Builder = 300 req/min).
How do I know the docs aren't stale? CI curls every endpoint these skills document against production nightly (scripts/check-endpoints.mjs). A renamed field or dead endpoint fails the build.
Can I use this commercially? MIT. Go ahead.
Reproducible, keyless, raw responses committed: the same 20 trader questions against CoinLobster's hosted MCP, CoinGecko's hosted MCP, and no tools. Run 2026-07-30: CoinLobster 14 yes / 5 partial / 1 no (sole source on 11 questions); CoinGecko 3 / 5 / 12 (and it cleanly wins price and volume breadth, which is its job); no tools 0 / 1 / 19. Failed calls are recorded as failed. See benchmarks/ for the scorecard, method and every raw call.
Issues and PRs welcome. The bar for changes: every claim in a SKILL.md must be checkable against a live endpoint, and no forecast language survives review. Run a skill's example curl calls before submitting.
The MCP server must stay declared in .mcp.json at the plugin root. Verified on Claude Code 2.1.220, for a marketplace plugin with "source": "./":
mcpServersdeclared inline in.claude-plugin/plugin.jsondoes not load.claude plugin detailsreports "MCP servers (0)".- A root
.mcp.jsondoes load.claude plugin detailsreports "MCP servers (1)" andclaude mcp listshows it Connected.
It looks like redundant config. It is not. Leave it.
- MCP server: coinlobster-mcp · hosted at
https://coinlobster.com/mcp - Agent-readable live pages: radar.md · crypto-whales.md · whale-statistics.md · llms.txt
- Connector setup (Claude, ChatGPT, Gemini): coinlobster.com/connect
MIT © CoinLobster