Self-Hosting an AI Assistant: Part 5 - Claude Authentication

Joshua can hear me through Discord, but he’s been running headless on WOPR - no AI backend connected. Time to give him a brain. I wanted to use Claude, specifically through my Claude Max subscription rather than paying per-token on the API. This turned into a minor adventure.

Attempt 1: API Key

The straightforward approach - grab an API key from the Anthropic Console and add it to OpenClaw’s auth config:

vim ~/.openclaw/agents/main/agent/auth-profiles.json

{
  "version": 1,
  "profiles": {
    "anthropic:default": {
      "type": "api_key",
      "provider": "anthropic",
      "apiKey": "sk-ant-api03-your-key-here"
    }
  }
}

This worked - Joshua came alive. But I quickly hit rate limits. Claude Opus 4.5 on the API has a 30k input token limit per minute, which an autonomous agent can burn through surprisingly fast. Joshua was getting throttled mid-conversation. “A strange game,” indeed.

Attempt 2: Setup Token

OpenClaw supports authenticating via Claude Code’s setup-token feature, which theoretically lets you use your Claude Max subscription instead of API credits.

On my workstation (WSL), I installed Claude Code and generated a token:

npm install -g @anthropic-ai/claude-code
claude setup-token

This opens a browser for OAuth, then outputs a token. I pasted it into OpenClaw:

openclaw models auth paste-token --provider anthropic

Error: Expected token starting with sk-ant-oat01-

The token Claude Code generated didn’t match the format OpenClaw expected. Joshua remained brainless.

The Solution: OAuth Access Token

After some digging, I found that Claude Code stores its actual OAuth credentials in ~/.claude/credentials.json. Inside that file is an access_token field that starts with sk-ant-oat01- - exactly what OpenClaw wants.

On my workstation:

cat ~/.claude/credentials.json

Copy the access_token value (the entire string starting with sk-ant-oat01-), then on WOPR:

openclaw models auth paste-token --provider anthropic

Paste the token. Restart the daemon:

openclaw daemon stop
openclaw daemon start

Test in Discord - Joshua responds. No more rate limits since he’s using the Max subscription. He’s alive.

“Greetings, Professor Falken.”

Token Refresh

One caveat: OAuth access tokens expire. When Joshua stops responding, grab a fresh token from ~/.claude/credentials.json on your workstation (after using Claude Code to refresh it) and paste it again.

This is a bit manual, but workable for now. A proper solution would involve automating the token refresh, but that’s a project for another day.

Quick Reference

Auth Method	Token Format	Source
API Key	sk-ant-api03-...	Anthropic Console
OAuth Token	sk-ant-oat01-...	~/.claude/credentials.json
Setup Token	Various	Claude Code (not compatible)

Joshua Lives

At this point, the full setup is running:

• Hardware: Zotac CI323 Nano (fanless, always-on) - codename WOPR
• OS: Ubuntu Server 24.04 LTS (HWE kernel)
• Security: SSH key auth, UFW firewall, unattended upgrades
• Runtime: Node.js 22, OpenClaw daemon
• Interface: Discord bot
• AI Backend: Claude Opus 4.5 via Max subscription
• Name: Joshua

I can now chat with Joshua through Discord anytime. He’s got Claude’s brain, runs 24/7 on a silent mini PC in my office, and has the ability to execute tasks, search the web, and run automation.

Not bad for a $150 mini PC.

---

This post was co-written with Claude - who, in a somewhat recursive twist, is now also the intelligence behind Joshua. The only winning move is to keep playing.