Use Claude Code with abliteration.ai

Route Anthropic’s Claude Code CLI through abliteration.ai and configure the correct context window for each Abliteration model.

To use Claude Code with abliteration.ai, set ANTHROPIC_BASE_URL to https://api.abliteration.ai and ANTHROPIC_AUTH_TOKEN to your ak_... key, then map Claude Code to an Abliteration model. Large V2 is the recommended text-only model for coding and has a 1M-token context window.

Setup is three short steps: set the auth variables, choose a model, then persist your config so you don’t re-export it in every terminal.

Install

macOS / Linux

curl -fsSL https://claude.ai/install.sh | bash

Step 1 — Set the auth variables

These two variables route Claude Code to abliteration.ai and authenticate the session. Every setup below needs them.

Variable	Value	What it does
ANTHROPIC_BASE_URL	https://api.abliteration.ai	Routes Claude Code to abliteration.ai instead of api.anthropic.com.
ANTHROPIC_AUTH_TOKEN	ak_YOUR_API_KEY	Recommended. Sends your abliteration.ai key as the bearer Authorization header. ANTHROPIC_API_KEY=ak_... also works (sent as x-api-key); pick whichever your environment already has wired.

Step 2 — Choose a model

Claude Code’s model picker is built around Claude model IDs. Use the following Claude Code values for Abliteration models:

Model	Claude Code value	Context	Input
Large V2	abliterated-model-large-v2[1m]	1,000,000	Text
Base	abliterated-model	262,144	Text and images
Previous large	abliterated-model-large[1m]	1,000,000	Text

Claude Code 2.1.223 and later assumes that unrecognized third-party model IDs have a 200K context window. The [1m] suffix tells Claude Code to use a 1M window. Claude Code removes the suffix before the API request, so abliteration.ai receives the canonical model ID, such as abliterated-model-large-v2.

Replace the default Opus / Sonnet / Haiku models

Override the model ID Claude Code sends for each default tier so the existing picker uses Large V2. Plain claude then launches into Large V2.

export ANTHROPIC_DEFAULT_OPUS_MODEL="abliterated-model-large-v2[1m]"
export ANTHROPIC_DEFAULT_SONNET_MODEL="abliterated-model-large-v2[1m]"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="abliterated-model-large-v2[1m]"

claude

Add Large V2 to the picker

Keep the default Claude tiers intact and add Large V2 as an extra picker entry you switch to explicitly.

export ANTHROPIC_CUSTOM_MODEL_OPTION="abliterated-model-large-v2[1m]"
export ANTHROPIC_CUSTOM_MODEL_OPTION_NAME="Abliterated Large V2"

claude --model 'abliterated-model-large-v2[1m]'

You can also switch to it inside a session with /model.

Step 3 — Persist your config

Exporting variables in every new terminal gets old. Pick one place to save them so Claude Code always picks them up.

Location	Where	Use when
Claude Code settings	~/.claude/settings.json	You want the config scoped to Claude Code only.
Shell profile	~/.zshrc or ~/.bashrc	You want the variables available to any Anthropic-compatible tool you run from the terminal.

Option A — ~/.claude/settings.json

Claude Code reads an env block from its own settings file. Values here apply whenever you start claude.

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.abliteration.ai",
    "ANTHROPIC_AUTH_TOKEN": "ak_YOUR_API_KEY",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "abliterated-model-large-v2[1m]",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "abliterated-model-large-v2[1m]",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "abliterated-model-large-v2[1m]"
  }
}

Option B — Shell profile

Append the same values to your shell profile so they export on every terminal launch.

# Append to ~/.zshrc, ~/.bashrc, or ~/.profile
export ANTHROPIC_BASE_URL="https://api.abliteration.ai"
export ANTHROPIC_AUTH_TOKEN="ak_YOUR_API_KEY"
export ANTHROPIC_DEFAULT_OPUS_MODEL="abliterated-model-large-v2[1m]"
export ANTHROPIC_DEFAULT_SONNET_MODEL="abliterated-model-large-v2[1m]"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="abliterated-model-large-v2[1m]"

# Reload
source ~/.zshrc   # or source ~/.bashrc

Context windows and compaction

Claude Code automatically compacts long sessions by summarizing earlier turns while keeping recent context. abliteration.ai does not expose a separate compaction endpoint; Claude Code sends the summarization request through the same ANTHROPIC_BASE_URL and model configuration.

For Large V2 and the previous large model, use the [1m] model values shown above. If you configure one model with its canonical ID instead, set its native context window explicitly. For example, this single-model configuration also enables the full Large V2 window:

{
  "env": {
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "abliterated-model-large-v2",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "abliterated-model-large-v2",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "abliterated-model-large-v2",
    "CLAUDE_CODE_MAX_CONTEXT_TOKENS": "1000000"
  }
}

CLAUDE_CODE_MAX_CONTEXT_TOKENS applies to the entire Claude Code process. Set it to 262144 when a configuration uses only abliterated-model. Prefer the model-specific [1m] suffix for the large models when you switch between models with different context windows.

CLAUDE_CODE_AUTO_COMPACT_WINDOW sets an optional, smaller compaction threshold. It does not declare a model’s native context length. Omit it unless you intentionally want Claude Code to compact before the native limit.

Headless mode

For CI/CD or scripts:

claude --model 'abliterated-model-large-v2[1m]' -p "summarize README.md"

Notes

• Published policies scoped to the key’s project apply to every Claude Code turn. See Policies.
• Large V2 and the previous large model are text-only. Use abliterated-model when a task includes images. See Models.