Use uncensored models in Claude Code

Claude Code talks to Silk Compute through the Anthropic Messages endpoint. Set these variables, then start claude.

Setup

Shell

export ANTHROPIC_BASE_URL=https://api.silkcompute.com
export ANTHROPIC_AUTH_TOKEN=$SILK_API_KEY
export ANTHROPIC_MODEL=deepseek-v4.1-flash-1017-uncensored
export ANTHROPIC_DEFAULT_OPUS_MODEL=deepseek-v4.1-flash-1017-uncensored
export ANTHROPIC_DEFAULT_SONNET_MODEL=deepseek-v4.1-flash-1017-uncensored
export ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek-v4.1-flash-1017-uncensored
export CLAUDE_CODE_SUBAGENT_MODEL=deepseek-v4.1-flash-1017-uncensored
export CLAUDE_CODE_MAX_CONTEXT_TOKENS=1048576
claude

To keep them, add the same variables to the env block of ~/.claude/settings.json.

Why every variable

Claude Code asks for its own model names for subagents and background tasks, so every model variable points at a Silk Compute model. CLAUDE_CODE_MAX_CONTEXT_TOKENS tells it the real context window, so it compacts at 1M tokens instead of 200K. Anthropic server tools such as web search are not available.

Subagents

Your account runs one request at a time, so subagents that run in parallel wait for each other. See Limits and errors.