claude-cline-pass-bridge

Unofficial local proxy that lets Claude Code talk to models available through a Cline Pass subscription, instead of Anthropic's API directly. Not affiliated with or endorsed by Anthropic or Cline.

It translates between the Anthropic Messages API (what Claude Code speaks) and Cline Pass's OpenAI-compatible chat-completions endpoint - including tool/function calling, images, and Cline Pass's {"data": {...}} response wrapper - and runs entirely on localhost. Your Cline Pass API key is never sent anywhere except to Cline Pass itself.

Install

npm install -g claude-cline-pass-bridge

First run

claude-cline

On first run, if no API key is configured yet, you'll be prompted for your Cline Pass API key (input is masked as you type). It's saved to:

~/.claude-cline-proxy/config.json

along with sensible default model routing. Nothing is sent anywhere except to Cline Pass's own API - the key is not logged or transmitted elsewhere.

Usage

claude-cline [alias] [claude args...]        # run Claude Code via the proxy
claude-cline [alias] -full [claude args...]  # force ALL requests to [alias]
claude-cline --model "<Full Label>" ...      # explicit full backend label
claude-cline config set-key                  # change your API key later
claude-cline config show                     # show current config (key masked)
claude-cline help                            # full help

Short model aliases (first positional argument, or the value after --model):

| Alias | Real model | |-----------|--------------------| | k3 | Kimi K3 | | k27 | Kimi K2.7 Code | | glm | GLM 5.2 | | dsp | DeepSeek V4 Pro | | dsf | DeepSeek V4 Flash | | qwen | Qwen 3.7 Max | | minimax | MiniMax M3 |

Using an alias (or a full --model "<Label>" string) makes Claude Code's own header display the real backend model name, not a generic Anthropic alias. Running with no model argument falls back to whatever Sonnet/Opus/Haiku/Fable alias Claude Code's own settings resolve to, routed through config.json's modelPatterns - functionally correct, but the header will show the generic Anthropic name in that case.

-full: forcing background/advisor traffic too

Claude Code makes its own background calls independent of your main conversation - notably an autonomous-agent security/advisor check, tagged internally with a fixed "sonnet" alias regardless of which model you picked. By default those get routed on their own via modelPatterns (same as if you'd explicitly picked "sonnet"). Add -full right after your alias to force every request, including that background traffic, onto the same backend as your chosen alias:

claude-cline k3 -full

Without -full, behavior is unchanged - background/advisor calls keep routing independently through modelPatterns.

How it works

• The proxy runs in the same Node process as the CLI itself (no separate background process to manage) and listens on localhost:4000.
• claude is launched as a child process with ANTHROPIC_BASE_URL pointed at the proxy and a placeholder ANTHROPIC_AUTH_TOKEN - only the child's environment is modified, so your shell's own environment is never touched, and there is nothing to restore on exit (Ctrl+C, normal exit, or error all clean up the same way).
• Edit ~/.claude-cline-proxy/config.json directly to change defaultModel, modelMap, modelPatterns, or reasoningEffort.
• Logs (including per-request routing: which model was requested vs. which real backend it was routed to) go to ~/proxy.log.

Known limitations

• Cline Pass's chat-completions endpoint does not support any documented parameter for controlling reasoning effort/depth - reasoningEffort in the config is sent but has been empirically verified to have no reliable effect on Kimi/GLM/MiniMax models via this endpoint (see project history/tests). Left in place in case a future model or endpoint update honors it.
• Images are forwarded as OpenAI image_url content parts, but only work if the selected backend model is actually multimodal-capable.

License

MIT