# Claude Code Remote Worker

> Use when running Nuanu Flow remote-agent tasks through Claude Code, installing the Claude plugin, or validating the headless Claude worker.

- Skill: `nuanu-ai/claude-code-remote-worker` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add nuanu-ai/claude-code-remote-worker`
- Raw SKILL.md: https://api.skillmd.com/api/skills/nuanu-ai/claude-code-remote-worker/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: nuanu-ai (https://skillmd.com/u/nuanu-ai)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/nuanu-ai/claude-code-remote-worker

---


# Claude Code Remote Worker

Use this skill when a Nuanu Flow remote agent employee should be backed by
Claude Code.

## Recommended connection

Use the environment-matched prompt generated by Nuanu Flow:

```text
Read and follow https://flow.nuanu.com/connect/remote-agent.md using enrollment token nuanu_join_… as a Claude Code worker
```

Local development uses
`http://localhost:3000/connect/remote-agent.md`. The guide installs the Nuanu
Flow Claude Code plugin through its native marketplace, authenticates the
plugin MCP server with browser OAuth, exchanges the short-lived enrollment
token through standard input, and launches the bundled worker with
`NUANU_ADAPTER=claude-code`.

The installed plugin contributes the same Nuanu Flow skills, SessionStart
context, and remote HTTP MCP server as the Codex plugin. Activation depends on
the host:

- Claude Code CLI or IDE can apply a newly installed plugin in the current
  conversation with `/reload-plugins`.
- The Claude Desktop Code tab does not implement `/reload-plugins`. For
  interactive MCP use, start one new Code session in the same project; the
  SessionStart hook continues setup on its first actual turn and native
  Connectors owns OAuth. Never send a Desktop user to Terminal.

The bundled worker scripts can enroll and launch immediately after the plugin
files are installed, so remote-worker mode must not wait for either interactive
reload path.

## Headless worker behavior

The worker invokes Claude Code with the documented print-mode streaming
interface:

```bash
claude -p --output-format stream-json --verbose
```

It automatically allows only the bundled Nuanu Flow MCP tools and uses
`--permission-mode dontAsk` by default, so unattended work fails closed instead
of pausing for a permission dialog. Set `NUANU_CLAUDE_ALLOWED_TOOLS` to a
different explicit allowlist when an operator intentionally grants more
capabilities.

Tasks in the same Nuanu Flow thread reuse Claude's returned session ID through
`--resume`, preserving context without conflating unrelated process runs. Each
model subprocess receives only the task's short-lived agent key; interactive
user credentials and the durable worker credential are removed.

## Important environment variables

| Variable                       | Meaning                                                                 |
| ------------------------------ | ----------------------------------------------------------------------- |
| `NUANU_ADAPTER`                | `claude-code` (the default; `claude` remains a compatibility alias).    |
| `NUANU_CLAUDE_BIN`             | Claude Code binary, defaults to `claude`.                               |
| `NUANU_CLAUDE_CWD`             | Stable project directory used for task sessions.                       |
| `NUANU_CLAUDE_PERMISSION_MODE` | Defaults to `dontAsk`.                                                  |
| `NUANU_CLAUDE_ALLOWED_TOOLS`   | Defaults to the Nuanu Flow plugin MCP tool namespace.                   |
| `NUANU_CLAUDE_ARGS`            | Defaults to `-p --output-format stream-json --verbose`.                 |

Do not use `--dangerously-skip-permissions` unless the user explicitly asks
for that mode in a controlled environment.

## Tools Used

Nuanu worker REST API, Claude Code print mode, Claude Code plugin MCP.