nc://claude-code · Nodecode

nodecode-claude-code

A Nodecode add-on that runs Claude models through your own logged-in Claude Code CLI. Pick claude-code/<model> in /models, and each model round goes out with the login claude already has.

Nodecode keeps the agent. Its turn loop, tools, context engine, retries and usage record work the same as on any other provider. The CLI only writes the request:

Nodecode builds the round's Messages request, the same way it does for Anthropic.
The add-on starts claude -p in stream-json mode with its own tools off. It hands the CLI the round's system prompt, history and tools, and points ANTHROPIC_BASE_URL at a loopback relay only that run knows.
The CLI writes the request it would make, with its own headers. The relay keeps it and the CLI is stopped, so the CLI itself never sends anything to Anthropic.
Nodecode sends those exact bytes to api.anthropic.com and reads the stream with its own Anthropic parser. Tool calls come back under Nodecode's tool names and run in Nodecode, never in the CLI.

It follows NousResearch's hermes-plugin-claude-subscription-directsdk.

Before you use it

Anthropic's terms. The Claude Agent SDK documentation says third-party products may not offer claude.ai login or plan rate limits without Anthropic's approval. This add-on is for running your own CLI on your own machine. Read the current terms yourself before relying on it.
Billing. Whether a round draws from your plan or from extra usage is decided by Anthropic's service. Turn off extra usage in your account if you don't want overage charges. Nodecode puts no price on claude-code rounds: none is billed per token, so the Dashboard shows no cost for them.
Models. Models with windows over 200K run on their 1M route. On some plans a 1M route, and Fable, draw from usage credits.

Install

Linux and macOS.

Install Claude Code, run claude once and /login. Then copy this folder into ~/.nodecode/addons/nodecode-claude-code/ and restart Nodecode (or run (restart-addons)). When no claude is on PATH, a notice says so.

Configure

This is optional. The section is on unless it says "enabled": false:

{
  "claude-code": {
    "command": "claude",
    "models": ["claude-sonnet-5", "claude-haiku-4-5"],
    "timeout_seconds": 120
  }
}

See config.example.jsonc.

What it costs in time

Every round starts a fresh claude process, about 0.4 s. The CLI then replays the session's history, about 20 ms for each earlier user turn. The CLI adds its own per-request context to the newest turn: the date, the account's email, and a short environment note naming the session's folder. The add-on moves the CLI's cache marker onto the part of the history the next round sends again unchanged, so earlier rounds are still read from cache.

Test

From a Nodecode checkout's src/:

sbcl --non-interactive --eval '(require :asdf)' \
  --eval '(push (truename ".") asdf:*central-registry*)' \
  --eval '(push #p"/path/to/addons/nodecode-claude-code/" asdf:*central-registry*)' \
  --eval '(asdf:test-system :nodecode-claude-code)'

The tests never start the real CLI and never reach Anthropic. A whole round runs against test/fake-claude.py, a stand-in that speaks the CLI's stream-json and needs python3.

MIT licensed.