How CLI Workers Work

A CLI worker spawns the Claude Code CLI as a subprocess and routes chat completion requests through it. From the API caller’s perspective, a CLI worker is indistinguishable from a GPU worker — the same OpenAI/Anthropic-compatible endpoints work with both.

This is pass-through inference: the router sends the request, the worker relays it to Claude Code, and streams tokens back. There is no agent framework, no skill injection, and no context file management — that lives in the commercial server.


Prerequisites

npm install -g @anthropic-ai/claude-code
claude login   # Interactive OAuth — or use CLAUDE_CODE_OAUTH_TOKEN for headless

Running a CLI Worker

CLI workers use the same auxot-worker binary as GPU workers. The router’s policy determines which mode a connected worker runs in — GPU or CLI. To run a CLI worker, connect auxot-worker with a key that the router has assigned CLI mode:

AUXOT_GPU_KEY=adm_xxx \
  AUXOT_ROUTER_URL=router:8080 \
  ANTHROPIC_API_KEY=sk-ant-... \
  ./auxot-worker

Or with Docker:

docker run -d \
  -e AUXOT_GPU_KEY=adm_xxx \
  -e AUXOT_ROUTER_URL=router:8080 \
  -e ANTHROPIC_API_KEY=sk-ant-... \
  ghcr.io/auxothq/auxot-worker:latest

The worker connects to the router, fetches its policy, and starts handling jobs using Claude Code as the inference backend. Ensure Claude Code is authenticated — set CLAUDE_CODE_OAUTH_TOKEN for headless deployments.


Authentication

MethodHow
InteractiveRun claude login once on the machine; credentials cached in ~/.claude/
HeadlessSet CLAUDE_CODE_OAUTH_TOKEN=sk-ant-... in the environment

Model Selection

VariableValuesDescription
AUXOT_CLI_DEFAULT_MODELsonnet, opusDefault Claude model to use

Callers can also specify the model per-request via the model field in the API request.


Tool Control

SettingValueEffect
AUXOT_CLI_TOOLS(unset)No tools enabled (default)
AUXOT_CLI_TOOLS"Bash,Read,Write"Only the listed tools
AUXOT_CLI_TOOLS"All"All tools enabled (use with caution)

Economics

A Claude Pro or Max subscription gives unlimited inference through the CLI worker at a fixed monthly cost — see anthropic.com for current pricing. For high-volume workloads, compare this against per-token API pricing.


Difference from the Commercial Server

The commercial Auxot Server also uses CLI workers, but wraps them in a full agent framework with skill injection, context files, and Tool Worker Policies — where MCP packages and credentials are configured per policy and dispatched through connected tools workers. If you need those features, see Auxot Server →.