defi-skills
Beta: This project is under active development.
Translate natural language into unsigned DeFi transaction payloads. A data-driven playbook engine resolves human-readable parameters (token symbols, ENS names, decimal amounts) into ABI-encoded calldata, with zero protocol-specific code in the engine.
13 protocols. 53 actions. All driven by JSON playbooks.
https://github.com/user-attachments/assets/cdc2fd63-b007-40a4-900e-a6f775b6e9fa
Quick Start
Install via pip (recommended)
pip install defi-skills --extra-index-url https://nethermind.jfrog.io/artifactory/api/pypi/kyoto-pypi-local-prod/simple
Or clone the repo
git clone https://github.com/NethermindEth/defi-skills
cd defi-skills
python3 -m venv .venv && source .venv/bin/activate
pip install -e .
Get started
defi-skills config setup # interactive wizard
defi-skills actions # list all supported actions
Deterministic mode (no LLM)
These examples work with no API keys at all (tokens resolved from local cache):
defi-skills build --action aave_supply --args '{"asset":"USDC","amount":"500"}'
defi-skills build --action lido_stake --args '{"amount":"5"}' --json
defi-skills build --action weth_wrap --args '{"amount":"2"}' --json
Actions that need on-chain data (swaps, "max" amounts, ENS names) require ALCHEMY_API_KEY:
defi-skills build --action uniswap_swap --args '{"asset_in":"WETH","asset_out":"USDC","amount":"0.5"}' --json
Interactive chat mode (LLM-powered agent)
defi-skills chat
defi-skills chat --model gpt-5.1
defi-skills chat --stream # raw text streaming instead of formatted output
An interactive agent that understands natural language, plans multi-step operations, and calls the deterministic engine via tool use. It discovers actions, checks parameters, and builds transactions. Type /help inside the chat for commands.
By default, responses are rendered with formatted Markdown after the LLM finishes. Use --stream for raw real-time text output, or persist the preference with defi-skills config set stream_responses true.
Simulation (WIP)
Transaction simulation (fork mainnet via Anvil, execute, report gas/transfers/balance changes) is implemented and available for manual testing via defi-skills simulate, but is not yet exposed to the chat agent or skill files. This is actively being worked on.
Agent Integration
This tool is built to be used by AI agents. It ships as a skill that any agent with shell access can pick up and use to build DeFi transactions deterministically, with no LLM running inside the CLI.
How it works for agents
- The user tells the agent something like "Supply 100 USDC to Aave"
- The agent (Claude Code, OpenClaw, or any LLM with shell access) classifies the intent and picks the right action + arguments
- The agent calls the CLI with
--action and --args (deterministic, no API key needed)
- The CLI returns unsigned transactions as JSON
- The agent presents the result to the user for review and signing
The agent is the LLM. The CLI is the execution engine. No LLM runs inside the CLI in this flow.
Skill files
Ready-made skill files are included for both Claude Code and OpenClaw:
| Agent | Skill location |
|---|
| Claude Code | .claude/skills/intent-to-transaction/SKILL.md |
| OpenClaw / generic agents | openClaw-Skill/SKILL.md |
Both contain the same instructions. Copy the appropriate file into your project and ensure defi-skills is installed with WALLET_ADDRESS set. The agent can then invoke it naturally:
"Supply 100 USDC to Aave"
"Swap 0.5 ETH for USDC on Uniswap"
"Buy PT for wstETH on Pendle"
The agent workflow is always: discover actions (defi-skills actions --json), check parameters (defi-skills actions <name> --json), build transaction (defi-skills build --action <name> --args '{...}' --json).
Claude Code Plugin
If you use Claude Code, you can install the skill directly as a plugin:
/plugin marketplace add NethermindEth/defi-skills
/plugin install defi-skills@nethermind-defi-skills
After installing, invoke it with:
/intent-to-transaction
Why deterministic mode for agents?
Agents already have an LLM (themselves). Running a second LLM inside the CLI would be redundant, slower, and more expensive. The --action + --args path uses zero LLM tokens and gives agents full control over intent classification while the engine handles the hard parts: token resolution, decimal conversion, ABI encoding, and approval handling.
How It Works
The CLI has two modes. The build command is purely deterministic (no LLM). The chat command adds an interactive LLM agent on top.
NethermindEth