elytro-wrapper

A natural-language frontend for the Elytro Ethereum smart-account CLI. You describe what you want in plain English; elytro-wrapper translates it into the correct elytro command, executes it, and returns the result as structured JSON.

The model runs entirely on your machine — no remote API calls are made.

How it works

User input (natural language)
        │
        ▼  node-llama-cpp + Qwen3-0.6B (local)
elytro CLI command(s)
        │
        ▼  child_process.spawn
JSON result

Requirements

• Node.js ≥ 18
• @elytro/cli installed globally

npm install -g @elytro/cli

Installation

npm install -g @elytro/wrapper

The postinstall script downloads the Qwen3-0.6B-Q8_0.gguf model (~640 MB) automatically. If the download is interrupted you can re-run it at any time:

npm run download-model

Usage

elytro-wrapper "<natural language input>"

Examples

elytro-wrapper "check my balance"
elytro-wrapper "show all accounts"
elytro-wrapper "what chain am I on"
elytro-wrapper "switch to agent-primary"
elytro-wrapper "send 0.01 ETH to 0xRecipient"

Output format

Every invocation prints a JSON array to stdout. Each element corresponds to one executed command:

[
  {
    "cmd": "elytro query balance",
    "succ": true,
    "stdout": "{ \"success\": true, \"result\": { ... } }",
    "stderr": ""
  }
]

| Field | Type | Description | | -------- | --------- | ---------------------------------------- | | cmd | string | The full command string that was run | | succ | boolean | true when the exit code was 0 | | stdout | string | Raw standard output from the command | | stderr | string | Raw standard error output from the command |

Error cases

| Situation | cmd | succ | |---|---|---| | Input cannot be mapped to a command | __UNKNOWN__ | false | | Command not in allowlist | original string | false | | Command execution fails | original string | false |

Multi-step commands

Some operations require multiple commands to be run in sequence. elytro-wrapper handles this automatically:

# swap: quote is always fetched before sending
elytro-wrapper "swap 1 ETH for USDC"
# → elytro swap quote ...
# → elytro swap send ...

# tx send: simulation always runs before broadcast
elytro-wrapper "send 0.01 ETH to 0xRecipient"
# → elytro tx simulate ...
# → elytro tx send ...

Development

# install dependencies
npm install   # also downloads the model via postinstall

# run without building
npm run dev "check my balance"

# build
npm run build

# run built output
node dist/index.js "check my balance"

Project structure

├── src/
│   ├── index.ts       # CLI entry point
│   ├── llm.ts         # LLM inference (node-llama-cpp + Qwen3-0.6B)
│   ├── prompt.ts      # System prompt
│   ├── executor.ts    # Command execution
│   └── whitelist.ts   # Command allowlist guard
├── scripts/
│   ├── download-model.mjs   # Model download script (postinstall)
│   └── load-env.mjs         # Dev .env loader
├── models/            # Downloaded model files (gitignored)
└── tsup.config.ts

Security

All commands generated by the LLM are validated before execution. Commands that do not start with elytro are rejected with an error result and never spawned.

Write operations (fund transfers, security configuration changes, etc.) are generated by the LLM and included in the output like any other command. The calling process or agent is responsible for prompting the user for confirmation before acting on write operations.