Claude Code reference-only setup.

Anthropic does not support routing Claude Code to non-Claude models. This path is unofficial and reference-only. Even at the current 131,072-token envelope, Claude Code 2.1.261 default mode stops before making an API request; --bare reaches AblitAI but receives the disabled tool-call gate, so the standard Claude Code harness remains unsupported even though API tools are on. Do not use paid credits yet.

Before you start

• Use Claude Code 2.1.261 with Node.js 22 or newer; native Windows also requires Git for Windows.
• The reference model ID is qwen3.8-27b-abliterated.
• The current deployment exposes a 131,072-token total context; completion must fit in the remaining envelope.
• Thinking, images, audio, and hosted web search remain unavailable. API tools are on; the standard Claude Code harness remains unsupported/reference-only.

1. Install the tested version

node --version
npm --version
npm install --global @anthropic-ai/claude-code@2.1.261
claude --version

2. Download the session settings

The consumer environment subset contains no key. It disables thinking, automatic retries, automatic updates, telemetry, error reporting, and other nonessential traffic for this session.

PowerShell

$settings = Join-Path $HOME ".claude\ablitai.env"
New-Item -ItemType Directory -Force (Split-Path $settings) | Out-Null
if (Test-Path $settings) { throw "$settings exists; review it before replacing it." }
Invoke-WebRequest "https://ablitai.com/configs/claude-code.env.example" -OutFile $settings
Get-Content $settings | ForEach-Object {
  if ($_ -match '^([A-Z0-9_]+)=(.*)$') { Set-Item "Env:$($Matches[1])" $Matches[2] }
}

macOS or Linux

settings="$HOME/.claude/ablitai.env"
mkdir -p "$(dirname "$settings")"
test ! -e "$settings" || { echo "$settings exists; review it before replacing it." >&2; exit 1; }
curl --fail --proto '=https' --tlsv1.2 https://ablitai.com/configs/claude-code.env.example -o "$settings"
chmod 600 "$settings"
set -a; . "$settings"; set +a

3. Set the key and clear conflicting providers

PowerShell

"ANTHROPIC_API_KEY", "CLAUDE_CODE_OAUTH_TOKEN", "CLAUDE_CODE_USE_BEDROCK", "CLAUDE_CODE_USE_FOUNDRY", "CLAUDE_CODE_USE_VERTEX" | ForEach-Object { Remove-Item "Env:$_" -ErrorAction SilentlyContinue }
$secret = Read-Host "AblitAI API key" -AsSecureString
$env:ANTHROPIC_AUTH_TOKEN = [System.Net.NetworkCredential]::new("", $secret).Password
Remove-Variable secret

macOS or Linux

unset ANTHROPIC_API_KEY CLAUDE_CODE_OAUTH_TOKEN CLAUDE_CODE_USE_BEDROCK CLAUDE_CODE_USE_FOUNDRY CLAUDE_CODE_USE_VERTEX
printf 'AblitAI API key: ' >&2
ablit_stty="$(stty -g)"
trap 'stty "$ablit_stty"' EXIT HUP INT TERM
stty -echo
IFS= read -r ANTHROPIC_AUTH_TOKEN
stty "$ablit_stty"
trap - EXIT HUP INT TERM
unset ablit_stty
printf '\n' >&2
export ANTHROPIC_AUTH_TOKEN

4. Confirm account access

Create or import your no-email account and save its offline backup. A valid key is required. This closed-tool diagnostic does not require funded credits and should return the documented 400 before inference or charging. Text inference separately requires a positive available balance.

5. Run the closed-gate probe

This --bare probe is expected to return 400 Tool calling is disabled for this deployment. Claude Code still includes tools when the prompt says not to use them. Reaching that exact error proves the reference environment reached AblitAI, but it is not a launch step and does not establish support.

claude --print --bare --model qwen3.8-27b-abliterated `
  --no-session-persistence `
  "Reply with exactly ABLITAI_OK. Do not use tools."

In every new terminal used for this diagnostic, reload the downloaded settings and prompt for ANTHROPIC_AUTH_TOKEN again. Do not start an interactive paid session unless this page is later promoted from reference-only to supported.

claude --print --bare --model qwen3.8-27b-abliterated \
  --no-session-persistence \
  'Reply with exactly ABLITAI_OK. Do not use tools.'

Why this diagnostic uses --bare

It avoids hooks, plugin sync, background prefetches, auto-memory, and project instruction discovery. A normal local request carried about 93 KiB of instructions and 26 tool definitions and currently returns Prompt is too long before reaching AblitAI; --bare carried about 4 KiB but still includes tools that Release A rejects. Large repositories and MCP catalogs further increase context use. A larger 32,768-token context is only a future candidate/revalidation plan. Never add --dangerously-skip-permissions; you should review and own every local tool permission.

Known client behavior

Claude Code labels the custom model as unrecognized and may display its own estimated dollar cost or model limits. Those client-side estimates are not AblitAI billing records. Your authoritative balance and token debit are shown on the account page.