Crypto DeFi Skills

Installable DeFi procedures for the agent you already use. Give Claude Code, Codex, Cursor, Gemini CLI, Hermes or OpenClaw a concrete task: inspect lending risk, review a swap, check a vault exit or trace a bridge delivery. 81 active skills across 17 independent packs. Named protocol skills contain the reads, calculations, decision branches and evidence needed to complete the task; official documentation supports those procedures.

Start with one skill

Install one workflow into your existing agent:

npx skills add galleonlabs/crypto-defi-skills --skill galleon-aave-position

Find your task · Install and run your first prompt · Agent index

Prefer a ready-made research desk to wiring skills into an agent? Boomkin is a local browser desk and installer built on these packs.

Do the skills change the answer? See the with/without benchmark: on 30 synthetic cases, 86 of 90 Claude Opus 5.5 answers with the skill and 66 of 90 without it were graded as meeting every expected criterion and avoiding every forbidden one. That historical benchmark does not measure the new protocol skills.

Choose a protocol task

Choose the skill that names the job. Each procedure includes the inputs to read, the decisions to make, failure handling and what counts as completion.

Each link opens an actual SKILL.md with bundled resources. Install its name with npx skills add galleonlabs/crypto-defi-skills --skill <name>. The task guide also covers market data, unsigned plans, security, payments and portfolio work. The agent index lists every active skill and its routing boundary. Two retired LP install names remain as compatibility notices.

Try a workflow

Replace the placeholders with your actual task inputs, then ask your agent:

Use galleon-aave-position to inspect [Ethereum wallet address]. Show the current health factor and a scenario with a 20% collateral-price fall and 30 days of debt interest. Keep the plan unsigned.

Use lifi-cross-chain to reconcile [source transaction hash] from Base to Arbitrum. Check [recipient address], the destination token and the amount actually received.

Use galleon-prediction-market-research to examine [Polymarket market URL]. Explain the exact resolution rules, outcome identities and available order-book depth. Keep this research-only.

For more examples, open a pack's README from the directory below and use its first prompt. Provider credentials stay in your agent's private environment or secret settings.

Independent packs

Install only the packs you need. Each has its own npm version, CLI, plugin manifests and self-contained skill resources; no pack requires another Galleon pack at runtime.

Install only what you need

The Agent Skills installer lets you select the receiving agent. List available skills or install a complete pack:

npx skills add galleonlabs/crypto-defi-skills --list
npx skills add https://github.com/galleonlabs/crypto-defi-skills/tree/main/packages/lending

Source commands follow main. For reproducible use, select a reviewed source revision or use an exact npm release. Published packages include a local corpus CLI and an ESM SKILL_CATALOG export:

npx --package [email protected] defi-lending-skills catalog --json

Node 20+ runs the published corpus CLIs. The CLI reads the bundled skills; the installer or Boomkin places them in your agent's discovery directory. Provider tools and optional helpers have their own runtime and access requirements.

Claude Code can install packs through /plugin marketplace add galleonlabs/crypto-defi-skills, then /plugin install galleon-defi-lending-skills@galleon-defi. Gemini CLI installs a pack natively with gemini skills install https://github.com/galleonlabs/crypto-defi-skills.git --path packages/lending/skills. Each pack also carries a Codex plugin manifest. The separate research plugin has its own ZIP installation path.

See the getting-started guide for pack selection, plugins, pinned versions and first-use checks. Installation adds procedures; it does not authenticate a provider or grant wallet authority.

What the skills give you

• A usable procedure. Concrete reads, amount units, calculations, decision branches and worked examples for the task.
• Official tools in context. Maintained SDK, API, CLI and MCP recipes with access, cost and deployment limits. Broad primitive workflows complement the named protocol procedures.
• Evidence you can inspect. Freshness and coverage limits, payload and permission review, transaction or order identity, and the balance or position evidence needed to finish.

Local helpers perform calculations or bounded diagnostics. The package CLIs do not sign or submit transactions. Research, unsigned planning, wallet execution and settlement stay distinct; preserve the user's authorization and resolve missing terms before consequential actions.

Sources and verification

The 8 October ETH.sh review maps 58 directory resources, records fetched source documents and distinguishes discovered interfaces from unavailable or unverified ones. It informed the agent, prediction, Aave V4 and streaming workflows, plus provider updates across the collection.

Skill quality explains structural checks, clean installs and recorded output evaluations. Research records retain primary sources and capability limits. Passing a format check, loading a skill or running a simulation does not establish a security audit, provider access or a financial outcome.

Develop and release

bun install --frozen-lockfile
bun run check
bun run pack

Use Bun 1.3.14 for development. Run clean consumer smoke tests when package contents or installation change. Start with CONTRIBUTING.md; agents editing the repository should read AGENTS.md.

The root workspace is private. Packs publish independently under <npm-name>@<version> tags using RELEASING.md. Discovery releases supply individual skill archives from an exact committed tree. Migration notes preserve older LP names and repository paths.

Contribute and reuse

Clearer examples, source-backed corrections and reproducible integration failures are welcome. Read the contributor guide, or use private vulnerability reporting for a security issue.

MIT licensed, created by Andrew Wilkinson and Galleon Labs. Retain the copyright and permission notice when reusing copies or substantial portions; ATTRIBUTION.md explains credit and third-party notices.