# LLM Worker

> remote-shell, opencode, ssh, project-isolation

- Skill: `lost-rob0t/llm-worker` (Agent Skill, multi-file: 3 files)
- Install (CLI): `npx get-skill add lost-rob0t/llm-worker`
- Raw SKILL.md: https://api.skillmd.com/api/skills/lost-rob0t/llm-worker/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: lost-rob0t (https://skillmd.com/u/lost-rob0t)
- Updated: 2026-11-18
- Page: https://skillmd.com/skills/lost-rob0t/llm-worker

---


# LLM Worker

## Goal

Run agent shell work on a project-isolated remote worker instead of the local
machine.

## Input

Configure `LLM_WORKER_HOST` and `LLM_WORKER_PROJECT`. The remote identity is
`PROJECT@HOST`; its project workspace is `~/repo`.

## Output

Remote stdout/stderr and the remote exit status are preserved by
`scripts/llm-worker-exec`.

## Workflow

```sh
export LLM_WORKER_HOST=worker.example
export LLM_WORKER_PROJECT=my-project
scripts/llm-worker-exec -- 'git status --short'
```

For OpenCode, run `scripts/install-opencode-bash-tool`. It installs a custom
`bash` tool that intentionally shadows OpenCode's built-in bash tool and routes
commands through this skill.

Prefer running OpenCode entirely on the worker when local file-edit tools and
remote shell state would otherwise diverge.

## Rules

- Fail closed when host or project is missing.
- Never infer a remote identity from private infrastructure.
- Keep credentials in SSH agent/config or runtime secret stores, never the skill.
- The project user's home and `~/repo` are owned by that project identity.
- Use `nix-remote-build` for builds that must return Nix outputs locally.

