Building a crypto research agent in 10 minutes with Cline + FalsifyLab

opener

For three years I've run a fleet of trading bots. Most die. The ones still alive eat from a curated daily feed of Form 4 cluster buys, 8-K material filings, ETF flows, DeFi yields with emissions stripped, Polymarket whale positions, and live macro tape.

Last week I packaged the same feed as an MCP server so AI coding agents can pull from it directly. This post walks through wiring it into Cline and using the agent to do real research without a dashboard.

If you've ever wanted Claude or another agent to answer "what insider buying is happening in healthcare today" or "show me Hyperliquid vaults with sharp 30d drawdown but still positive return", this is the post.

Repo: github.com/FalsifyLab/falsifylab-alpha-mcp
Free tier: no signup, just install + call

1. The setup

Cline is an autonomous coding agent that runs inside VS Code. Like Claude Code, but with a different orchestration loop. The thing that makes Cline useful for research (not just coding) is its MCP support — you can plug in arbitrary tool servers and the agent will discover + use them.

What we'll wire up:

• Cline as the agent runtime
• FalsifyLab Alpha MCP as the data layer (8 finance tools)
• Optionally a Python script** the agent writes for itself based on the data it pulls

The whole install is one line:

pip install falsifylab-alpha-mcp

Then add it to Cline. Open Cline settings → MCP Servers → Add Server:

{
  "mcpServers": {
    "falsifylab-alpha": {
      "command": "python",
      "args": ["-m", "falsifylab_alpha_mcp"],
      "env": {}
    }
  }
}

Restart Cline. The 9 tools are now discovered:

• top_yield_farms — DeFi yields, emissions stripped
• hl_vault_leaderboard — Hyperliquid vault rankings
• insider_buy_clusters — Form 4 cluster buys
• sec8k_material_today — 8-K filings filtered
• macro_tape — equity/vol/rates/FX/commodities/crypto snapshot
• etf_flow_today — US spot crypto ETF aggregates
• active_airdrop_farms — airdrop yield-gap detection
• polymarket_whale_positions — top wallet positions
• confluence_today — cross-source signal alignment

The free tier requires no signup, no key, no email. The server auto-issues a tier-0 key on first call. Real data. 24h cached. 10 results per query. Designed to be useful, not a teaser.

2. First query — what's happening right now

Open a new Cline task. Type:

Use the macro_tape tool to check the current macro regime. Are we risk-on or risk-off? Tell me what's moving and what's quiet.

Cline calls the tool. Gets back something like:

SPX: $7400.96, +1.95% 1d, +0.49% 5d
VIX: 18.02, +3.62% 1d, +3.62% 5d
UST 10Y: 4.32%, -2 bps 1d
DXY: 99.4, -0.3% 1d
GOLD: $4721, +0.47% 1d
WTI: $102.41, +8.02% 1d
BTC: $80,793, +0.76% 1d
ETH: $2,285, -0.95% 1d

The agent then writes an interpretation. Real example from a session I ran while writing this post:

"Risk mixed, leaning hedging-on. Equities up but VIX also up — that's not a clean rally, it's a rally being met with put-buying. Crude +8% one-day is the dominant macro shock. UST10Y down 2bps suggests bond bid (flight to quality despite equity strength) — also confirms hedging. Crypto soft, ETH lagging BTC by 1.7%. Posture for me: don't chase index longs, lean fade-the-vol-spike via OTM SPY calls if VIX holds above 17 into close."

This is grounded interpretation, not vibes. Every claim references a number from the tool output. The agent doesn't hallucinate macro data because the data was just handed to it.

3. Stacking signals — the confluence_today tool

The most interesting tool is confluence_today. It returns tickers/assets where 2+ FalsifyLab signals align in the last 24h. Insider cluster + material 8-K filing on the same ticker. ETF inflow streak + macro regime supporting the direction. DeFi yield spike + airdrop activity + HL vault concentration on the same token.

Why this matters: every individual signal is noisy. But signals stacking on the same asset is rarer, and historically more predictive.

Try this in Cline:

Use confluence_today (kind=equity, min_signals=2) and tell me what's lighting up. For each ticker, summarize why it's flagged and what the bear case might be.

The agent calls the tool, gets back equity-side stacking. Example output (real, slightly redacted):

{
  "equity": [
    {
      "ticker": "DCO",
      "signal_count": 2,
      "signals": [
        {"kind": "form4_buy_cluster", "filer_count": 3, "total_buy_usd": 142000},
        {"kind": "sec8k_material", "items": ["2.02"], "link": "..."}
      ]
    }
  ]
}

Cline then writes:

"DCO (Ducommun, aerospace/defense contractor) is the only flagged equity. Three insiders bought a combined $142k in the last 24h while the company filed an 8-K under item 2.02 (preliminary earnings). That's the textbook bullish stack: insiders know something + earnings-related news drop within the same window. Bear case: $142k is small money — could be planned grants. Verify by checking if the buys are coded P (open-market purchase) vs A (award)."

The agent did three things automatically:

1. Called the tool
2. Interpreted the JSON
3. Pre-empted the obvious bear case (10b5-1 vs discretionary)

I didn't tell it to do step 3. It knew because Form 4 trading mechanics are in its training data — the tool gave it the structured handle to apply that knowledge.

4. Writing a research bot

The next move is to let the agent write code that calls the tools on a schedule. Example task:

Write a Python script that calls insider_buy_clusters and sec8k_material_today every hour, finds tickers that appear in both, and emails me a summary at the end of the day. Use the FL_API_KEY env var if set; otherwise free tier. Save the script as daily_alpha_brief.py.

Cline writes the script. Reads the MCP server's tool schemas via tools/list. Writes a small Python loop. Adds basic error handling. Drops it into the workspace.

A few notes on what makes this work:

• The MCP server's inputSchema for each tool is detailed enough that Cline writes correct argument types on first try
• The mechanism field on yield/airdrop responses tells the agent how to describe positions to humans
• Free-tier auto-key-issuance means the script works for any user who runs it without a setup hassle

Caveats:

• Free tier rate-limits at 60 req/hr. Pro tier removes the limit. For a daily cron at 1/hr, free tier is fine
• The tools don't include trade execution. They're read-only data. Execution is your problem — keep it that way
• If you're tempted to let the agent write a trading bot directly, don't. Backtest. The data is honest, but every trader thinks their signal works until it doesn't

5. Going Pro

If you want real-time data instead of 24h cache, 100 results per query, and 90 days of history, FalsifyLab Pro is $19/mo with a 7-day trial. There's also a Pro Plus tier ($49/mo) with email + Slack alerts when confluence_today returns a new ticker — useful for agents that should fire on signal change rather than polling on a schedule.

Get an API key from falsifylab.com/pro, set FL_API_KEY=fl_yourkey in the Cline MCP env, restart Cline.

Launch promo: EARLY50 = 50% off first 3 months. Capped at 25 across all tiers.

There's also a free Pro Plus alternative: keep using the free tier and have Cline call the tools on a 15-minute cron. The data is 24h cached on free, so you won't catch real-time triggers, but for end-of-day research it works.

6. Closing

Agent-fit data is a different product than dashboard data. Token Terminal, Nansen, Dune are all built for humans clicking through UI. FalsifyLab is built for agents calling tools. Different latency, different schema, different price point.

If you're building an autonomous research workflow — for crypto, equity, or both — try the free tier and tell me what's missing. Issues at github.com/FalsifyLab/falsifylab-alpha-mcp/issues.

Live demo agent: falsifylab.com/demo (reads 8 tools every 15min, posts a memo).
Docs: github.com/FalsifyLab/falsifylab-alpha-mcp#readme