Use an uncensored model in Claude Code
Claude Code talks to the Anthropic Messages API, and Obscurify.ai serves that API at https://obscurify.ai/v1/messages. Point Claude Code at it and it runs on any of our models with tool support: uncensored models that don’t refuse, cybersecurity-tuned models for authorized security work, or Claude itself through an account with no email.
Tested on 2026-11-08 with Claude Code 2.1.319 against the live API: on OrcaCyber MiniMax M3.1 Flash and CyberHeabsy, Claude Code wrote a script with its Write tool, ran it with Bash and reported the output. During development the same flow passed on DeepSeek V4.1 Flash Uncensored and its thinking variant.
Before you start
- An Obscurify.ai account. It’s a random 16-digit number, with no email, phone or password. Sign up, then add funds with Monero, Bitcoin, Lightning or a card.
- An API key from the account page.
- Claude Code installed (
npm install -g @anthropic-ai/claude-codeor the native installer).
1. Pick a model
Claude Code needs a model that supports tool calling. In the API model list those have "tools": true. Good choices:
| Model id | Why | Plan |
|---|---|---|
deepseek/deepseek-v4.1-flash-uncensored | Uncensored, 1M context, fast, cheap | Pro |
z-ai/glm-5.3-flash-uncensored | Cheapest uncensored model with tools | Pro |
orca/orcacyber-minimax-m3.1-flash | Tuned for security work | Any account |
heabsy/cyberheabsy | Security-tuned, 1M context | Any account |
pqhaz/deepseek-v4.1-flash-1017-heretic | Decensored DeepSeek, cheapest with clean tool calls | Any account |
Uncensored models need the Pro plan. The cybersecurity and decensored models work on every plan, including Pay As You Go with no monthly fee. See which uncensored models refuse least and how the security models compare.
2. Set three environment variables
export ANTHROPIC_BASE_URL=https://obscurify.ai
export ANTHROPIC_AUTH_TOKEN=sk-obs1_... # your Obscurify.ai API key
export ANTHROPIC_MODEL=deepseek/deepseek-v4.1-flash-uncensored
Claude Code also uses a small, fast model for background jobs such as naming the session. Point it at the same model, or it asks for a Claude model:
export ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek/deepseek-v4.1-flash-uncensored
(Older Claude Code versions call this one ANTHROPIC_SMALL_FAST_MODEL.)
3. Tell Claude Code the context window
Claude Code doesn’t know our model ids, so it warns that it “assumes a context window of 200k tokens” and compacts the conversation there. Give it the real window from the model list:
export CLAUDE_CODE_MAX_CONTEXT_TOKENS=1000000 # 1M models; 260000 for OrcaCyber MiniMax M3.1 Flash
4. Run it
mkdir demo && cd demo
claude
Or one shot, the way we tested it:
claude -p "Create hello.py that prints hi, then run it with python3 and tell me the output." \
--allowedTools "Write" "Bash(python3:*)"
It writes the file, runs it and answers “hi”. You pay each model’s own price per token from your balance; there’s no Claude subscription involved.
Using Claude models this way
Bare Claude names work too. claude --model claude-sonnet-4-5 maps to anthropic/claude-sonnet-4.5 on Obscurify.ai, billed at that model’s price. We never swap in a different model: if we don’t carry the one you name, you get an error that says to set ANTHROPIC_MODEL. Claude still refuses what Claude refuses; the difference is that the account is a number, not your email.
Problems
- 401 authentication_error: the key is wrong or missing. Use
ANTHROPIC_AUTH_TOKEN(orANTHROPIC_API_KEY) with the key from the account page. - 402 billing_error: the balance is empty. Add funds on the account page.
- 403 permission_error, “not on plan”: an uncensored model on a plan below Pro. Pick a cybersecurity or decensored model, or upgrade.
- 404 not_found_error: the model id is wrong. Keep the maker prefix (
deepseek/...,orca/...). - Tool calls fail or the model just talks: the model doesn’t support tools. Pick one with
"tools": true. - Screenshots and large images are rejected: images over 8 KB aren’t accepted yet, so Claude Code’s image reads fail. Text, code and tools work.
What stays private
Prompts and replies aren’t stored on Obscurify.ai’s servers. Your identity is scrambled before a request reaches the model provider, so the provider sees the text, including the files Claude Code reads and sends, but not who sent it. Claude models don’t have the ZDR tag, so Anthropic’s own retention policy applies to them. Models that do have it only go to zero-data-retention routes; these rest on the providers’ published policies, not signed agreements. The account is a random number, and the API key is the only thing tying requests to it.
The terms apply to what you ask: no writing or spreading malware, and no breaking into systems you aren’t authorized to access.
More setups: uncensored models in coding agents · Cline · OpenCode · Aider · API docs