1.6.4
2026-08-30
Every data route audited by hand. Twelve added, 31 total. Correct auth headers, several endpoints that had quietly moved, and one route that was claiming more than it could answer.
Make your AI defi native: a true expert on onchain finance.
Your AI might know the yield. This skill makes it check who actually pays it. Get deep breakdowns on vault strategies, onchain trade types, and defi dynamics.
Works with Claude, ChatGPT, Grok Bot, Cursor, and any agent that reads Agent Skills.
121
sources, every URL checked
31
live data routes
119
terms defined
11
graded evals
See all
Lending
Morpho / Aave / Euler / Spark / Compound / Fluid / Silo / Gearbox / Kamino / Liquity V2 / Flex
Curators
Steakhouse / Gauntlet / Clearstar / MEV Capital / KPK / Sentora / Armitage / Birch Hill / TAU Labs
Vault infra
Upshift / Veda / Lagoon / Aera / Enzyme / Yearn / Mellow / Concrete / IPOR
Real world assets
Midas / Ondo / Securitize / Franklin Templeton / Superstate / Centrifuge / Grove / Plume
Stablecoins
Sky / Circle / Ethena / Tether / Resolv / Elixir / Usual / Lido
Markets
Pendle / Hyperliquid / Uniswap / Curve / Aerodrome / Derive / Lighter / Maple / 3Jane
Chains
Base / Solana / Monad / MegaETH / Robinhood Chain / Flashbots / Titan Builder
Oracles
Chainlink / Pyth / RedStone / Chronicle / SEDA
Data
DefiLlama / vaults.fyi / rwa.xyz / Merkl / Chaos Labs / LlamaRisk / Pharos
Academic
Loss-Versus-Rebalancing (Milionis et al.) / Pump-and-dump research (Kamps & Kleinberg; Dhawan & Putnins) / SEC microcap fraud alerts
+ many more. New data sources are continually added. Also covered: the ERC-4626 and ERC-7540 vault standards, restaking, bitcoin wrappers, tokenised equities and regulatory primaries.
01 The difference
Same question. One answer is useful.
AI is increasingly improving. It can do a lot without defi-native, but misses gaps. This skill makes it check multiple layers of the onchain finance stack.
Without the skill
>this vault pays 7% on USDC. worth putting $10k in?
7% is attractive compared to traditional savings rates. The vault uses established DeFi protocols. Yields can fluctuate, and smart contract risk always exists. Consider your risk tolerance and do your own research.
With defi-native
>this vault pays 7% on USDC. worth putting $10k in?
✓Names the five layers between you and the yield
✓Looks through to what is actually backing it, whether it is wrapped bitcoin or a synthetic dollar
✓Names the oracle, and whether liquidations can even fire
✓Breaks down what makes up yield
✓Understands risk tranches and who eats the first loss
✓Knows what PT, YT, and concentrated liquidity positions are
02 What you can do with it
Be confident ChatGPT or Claude is actually understanding defi.
Ask a question on crypto, defi, or onchain markets. Include links if you'd like. The skill loads itself after it's installed.
Understand what you're depositing into
Use defi-native to know what the vault holds, who set the parameters, what breaks first, if there's incentives, and where the yield comes from.
Assess this vault before I put $5k in
Break the APY apart
Who pays it, how much is real income versus token emissions, and what you actually keep after fees, denomination, and the size you are trying to fit.
Is this 12% sustainable, and what would I really earn?
Stay up to date onchain
Defi moves fast. A weekly read on rates, flows, and structure, plus a deep base on financial primitives to break down new onchain experiments.
What changed in DeFi this week, and what should I watch?
Learn true financial concepts
defi-native was built from 6,000+ pages on industry conferences, books on Wall Street, the history of money, protocol docs, and academic essays. Use it to learn how defi works and its similarities to traditional finance.
How do onchain options work? Build me a visual
03 How it works
Install once. Use ChatGPT and Claude as you do.
1
It stays out of the way
The skill wakes up when your question is actually about onchain yield, tokens, rates or market structure.
2
It routes to what it needs
Assessing a vault pulls a different file than explaining a covered call. Its deep reference material loads per task, never all at once.
3
It refuses to guess
Before any number goes in an answer it pulls a fresh one and dates it. It pulls from 30+ APIs.
See all data sources
Lending
Morpho / Aave / Euler / Spark / Compound / Fluid / Silo / Gearbox / Kamino / Liquity V2 / Flex
Curators
Steakhouse / Gauntlet / Clearstar / MEV Capital / KPK / Sentora / Armitage / Birch Hill / TAU Labs
Vault infra
Upshift / Veda / Lagoon / Aera / Enzyme / Yearn / Mellow / Concrete / IPOR
Real world assets
Midas / Ondo / Securitize / Franklin Templeton / Superstate / Centrifuge / Grove / Plume
Stablecoins
Sky / Circle / Ethena / Tether / Resolv / Elixir / Usual / Lido
Markets
Pendle / Hyperliquid / Uniswap / Curve / Aerodrome / Derive / Lighter / Maple / 3Jane
Chains
Base / Solana / Monad / MegaETH / Robinhood Chain / Flashbots / Titan Builder
Oracles
Chainlink / Pyth / RedStone / Chronicle / SEDA
Data
DefiLlama / vaults.fyi / rwa.xyz / Merkl / Chaos Labs / LlamaRisk / Pharos
Academic
Loss-Versus-Rebalancing (Milionis et al.) / Pump-and-dump research (Kamps & Kleinberg; Dhawan & Putnins) / SEC microcap fraud alerts
+ many more. New data sources are continually added. Also covered: the ERC-4626 and ERC-7540 vault standards, restaking, bitcoin wrappers, tokenised equities and regulatory primaries.
04 Install
One command. Sixty seconds.
No keys, no config, no account. Every core feature works without an API key.
$npx skills add emlai/defi-native-skill
?Which agent? detected automatically
?Install where? defaults are fine
✓defi-native v1.7.0 installed
$start a new session
1
Open a terminal, not a chat window
Mac: Cmd+Space, type Terminal, hit Enter. Windows: open PowerShell from the Start menu.
2
Paste the command
It finds your agent and asks where to put things. Take the defaults. If you see command not found: npx, install the LTS build of Node.js and run it again.
3
Start a new session and ask something
Try “what is actually backing sUSDe?” The skill loads on its own.
Easiest route if you are already in a coding agent. Paste this and it installs itself.
>Install the skill from github.com/emlai/defi-native-skill by running: npx skills add emlai/defi-native-skill
Works the same in Claude Code, Cursor, Codex, Grok Bot, or anything else that can run a shell command.
Clone it and link it in. This is also the install that can keep itself updated.
$git clone https://github.com/emlai/defi-native-skill.git ~/code/defi-native-skill
$ln -s ~/code/defi-native-skill ~/.agents/skills/defi-native
~/.agents/skills/ is the shared folder most hosts read. Claude Code also reads ~/.claude/skills/, so link it into both if you use several agents.
Optional API keys unlock deeper data and are never required. The skill will tell you which single key would improve the answer it is giving you, then carry on without it.
Detects your agent
Keyless by default
Read only, always
05 Staying current
Push to GitHub, and the skill updates itself.
A skill is files on disk, so by default it does not move when the repo does. Three ways to close that gap, from fully automatic to a one line nudge.
Live at the start of every session
Install the manual way, so the skill folder is a symlink to your clone. Now git pull is the update. Add a session hook that pulls before every run and the last thing you pushed is what your agent loads.
~/.claude/settings.json
Claude Code
{
"hooks": {
"SessionStart": [
{
"matcher": "startup",
"hooks": [
{
"type": "command",
"command": "git -C \"$HOME/code/defi-native-skill\" pull --ff-only -q || true"
}
]
}
]
}
}
--ff-only refuses to merge, so it can never touch a local edit. The || true means a session on a plane still starts. Skills load at session start, so a mid-session push lands in the next one.
For hosts without hooks
Same symlinked clone. Instead of a hook, pull on a timer so Cursor, Codex, Grok Bot and anything else reading that folder stay current too.
crontab -e
any host, mac or linux
# pull every 30 minutes, quietly, never merging
*/30 * * * * /usr/bin/git -C "$HOME/code/defi-native-skill" pull --ff-only -q
Or just let it tell you
On a plain install the skill checks itself. Once per session it compares its own version against the repo, and if it is behind it says so in one line, then keeps working. Installing is always your call, and the skill never rewrites its own files. A skill that silently edits itself cannot be reviewed.
defi-native 1.6.4 installed, 1.7.0 available.
$npx skills update
Make sure what you pushed is worth pulling
On every push the verifier re-checks the manifest schema, enforces the writing conventions, and fails if a reference file went missing from the README. A second check compares every number printed on this page against the skill itself, so the page cannot quietly outlive what it describes. Source liveness runs on a monthly schedule instead, so a vendor's docs going down never blocks a deploy.
.github/workflows/pages.yml
github actions
name: Publish site
on:
push:
branches: [main]
permissions:
contents: read
pages: write
id-token: write
jobs:
build:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- uses: actions/setup-python@v5
with:
python-version: "3.12"
# the same gate the pull request check runs
- run: python3 scripts/verify_manifest.py --skip-network
# refuse to publish a page whose numbers drifted
- run: python3 scripts/build_site.py
- uses: actions/upload-pages-artifact@v3
with:
path: site
deploy:
needs: build
runs-on: ubuntu-latest
environment: github-pages
steps:
- uses: actions/deploy-pages@v4
Switch it on once under Settings, Pages, Source, GitHub Actions. A monthly job already runs the same verifier and opens a pull request with proposed updates. Automation checks facts. People change judgment.
06 Changelog
Latest updates
The skill is updated based on output quality, feedback, and advances in defi.
1.7.0
2026-08-31
Market microstructure. Depth as a number instead of a vibe, the manipulation taxonomy with its onchain fingerprints, squeeze anatomy, and tokenized stocks read as three prices on two clocks. Plus attention assets and the amplification ratio, a seventh lending architecture where the borrower's rate is a bid for redemption-queue position, and a raw-log forensics recipe for windows the aggregators cannot see.
1.6.3
2026-08-30
A router for live data. Question patterns mapped to real endpoints, with the pitfalls attached and a note on what your own API key would unlock.
1.6.2
2026-08-30
No seat named without its numbers. Every alternative in a comparison now has to carry the same evidence as the main recommendation. An eval enforces it.
1.6.0
2026-08-30
Options, trade anatomy, curation. Every concentrated liquidity position read as the short option it is, the discounted locked-token trade taken apart with three verified 2026 prints, and eleven published curator frameworks distilled into one thing you can score a manager against.
1.5.3
2026-08-28
Four models tried to break it. Every confirmed finding fixed, including a clearinghouse waterfall in the wrong order and a price print that was off by 10x.
1.5.2
2026-08-28
Caught in the field. One query was returning only half of Morpho's vaults, so any agent following the old recipe silently missed every V2 book.
1.5.1
2026-08-28
First public release. Gated on paired runs with and without the skill, audited by independent models, every source verified live.
Full history in CHANGELOG.md, rebuilt here on every push.
07 Questions
Questions
What is an agent skill?
A folder of plain text files you drop into an AI agent to make it expert at one thing. Instructions, reference material, worked examples. The agent picks it up only when your request matches, so you get depth without stuffing every prompt.
Does it only work with Claude?
No. Any agent that reads the Agent Skills format: Claude Code, Cursor, Codex, Grok Bot and others. The installer detects your host. If you install by hand, the shared folder is ~/.agents/skills/.
Do I need API keys?
No. Everything core runs keyless. Keys buy you deeper data, and the skill will name the one key that would improve the answer in front of you rather than making you guess. It never waits on one.
Can it move my money?
No, and that is a hard rule in the skill, not a setting. It will not build, sign, submit or approve a transaction, and will not touch an allowance, no matter what tools are connected or how the request is worded. It shows you the action and hands it back.
The numbers inside will go stale. Does that break it?
That split is the whole design. The mental models age slowly and live in the reference files. Numbers age in weeks, so every figure in there is labelled a worked example and the skill pulls fresh data before it makes a numeric claim. Including about itself.
How much does it cost?
defi-native is completely free and always will be.
Is there a token?
The bankr community launched a $defi-native token based off this AI skill announcement on X. Fees are used to fund the project. Some fees have also been airdropped to holders. The defi-native team owns zero tokens.
Is this financial advice?
No. It produces research, and says so every time. Ask for a pick and you will get one, but only with the full view attached: what has to be true, what breaks it, the total loss case, and the runner up. DeFi has total loss tails and the skill never pretends otherwise.
Make your AI agent defi native today.
Free, MIT licensed, and it works in the agent you already use.
$npx skills add emlai/defi-native-skill