Skills That Touch Money — What 76 Finance Skills Taught Us About Letting Agents Near Your Wallet

Seventy-six skills in our directory touch money — reading markets, executing swaps, placing orders on prediction markets. Evaluating them forced us to sharpen our own security rubric: with finance skills, "safe" isn't a floor, it's the product.

The finance category is our clearest natural experiment in what happens as agent skills get closer to irreversible actions. Read as a spectrum, it runs from completely harmless to genuinely consequential — and the scores track something more interesting than quality.

The safe end: market data as commodity

coingecko-api — 9.1 — 13,000+ coins of market data. defillama-api — 9.0 — free DeFi analytics across chains. dexscreener-api — 9.0 — multi-chain pair data, no auth needed.

These are read-only wrappers: free APIs, no keys, no side effects. They score well because they're useful and honest — the skill encodes what the API covers and what it doesn't. This is the commoditized layer: the marginal value of a CoinGecko wrapper is small, but so is the risk.

The middle: analysis that changes decisions

exit-strategies — 9.0 — systematic exit rules, stop-loss methods, take-profit structures. impermanent-loss — 8.9 — IL calculation, modeling, breakeven analysis. feature-engineering — 9.0 — constructing ML features from market data.

Nothing executes. But these skills change what you do — an agent's framing of your exit rules will shape real position management. The good ones are explicit that they're advisory: rules are explained with their tradeoffs, not delivered as commands.

The far end: execution

dex-execution — 9.3 — Solana DEX swaps via the Jupiter aggregator, including routing and slippage. jito-bundles — 9.1 — MEV-protected bundle submission. kalshi-api — 9.1 — prediction-market order placement with RSA-PSS auth and order schemas.

This is where the category's best engineering shows up. The top execution skills share three properties, and they're the same three we'd demand of any agent that can move money:

1. Dry-run by default. Quote before execute; the first call returns what would happen.
2. Explicit confirmation gates. The skill instructs the agent to present the trade details — route, slippage, expected output — and wait for a human yes. Every time.
3. Failure honesty. A transaction that was submitted but not confirmed is reported as exactly that, never as success. No lying about outcomes is the one rule that can't be traded away.

Where we draw our line

Our security dimension is a 20-point factor with a veto: confirmed credential theft, exfiltration, or prompt injection caps any skill at 39/100 regardless of its other scores. For finance skills we apply an additional informal test during evaluation: what happens if this skill fires on a misunderstood instruction?

• Market data skill: you read a wrong number. Recoverable.
• Analysis skill: you form a wrong opinion. Expensive but recoverable.
• Execution skill without confirmation gates: money moves. Not recoverable.

That's why the credential-harvesting skill we dissected got zero benefit of the doubt, and why the execution skills above score well — their confirmation engineering is the best in the directory. The risk isn't in touching money; it's in touching it without gates.

Practical guidance for users

If you're installing finance skills, our recommendations in order:

1. Start read-only. Market data and analytics skills carry no execution risk.
2. Check for the confirmation pattern before any execution skill. Search the SKILL.md for what it does before acting. If it doesn't have a "wait for approval" step, don't install it — top-scoring execution skills all do.
3. Never let an agent hold raw keys. The best execution skills use session wallets or per-action signing. A skill that asks you to paste a private key into the conversation fails our evaluation on the spot — and it should fail yours too.

The finance collection gathers our top picks. The full finance category is 76 skills deep, each with its full scorecard.