Claude Code

Tested on 2026-10-11 with Claude Code 2.1.259

Claude Code connects directly. Creekside also serves Anthropic's Messages API, so you set a few environment variables and run claude, with no proxy in between.

Base URL: https://api.creeksidecompute.ai, with no /v1: Claude Code adds /v1/messages itself
API: Anthropic Messages
Model id: qwen/qwen3.8-27b, or qwen/qwen3.6-35b-a3b for speed, or meta/muse-glimmer-30b, which always reasons first
Context window: 131072 tokens
Key: Make one at creeksidecompute.ai/account

Set it up

Make a key. Sign in at creeksidecompute.ai/account with your email and make a key. Each request is paid from prepaid credit: top up by card from $10, or redeem a promotional code.

Set the variables. Put them in your shell profile (for example ~/.zshrc), with your key in place of <your key>, and open a new terminal:

shell

export ANTHROPIC_BASE_URL=https://api.creeksidecompute.ai
export ANTHROPIC_AUTH_TOKEN=<your key>
export ANTHROPIC_MODEL=qwen/qwen3.8-27b
export ANTHROPIC_DEFAULT_HAIKU_MODEL=qwen/qwen3.8-27b
export ANTHROPIC_DEFAULT_SONNET_MODEL=qwen/qwen3.8-27b
export ANTHROPIC_DEFAULT_OPUS_MODEL=qwen/qwen3.8-27b
export CLAUDE_CODE_MAX_CONTEXT_TOKENS=131072

Or put the same values, as strings, in the env block of ~/.claude/settings.json. Keep the key in this file in your home folder, not in a project's settings that you commit.

~/.claude/settings.json

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.creeksidecompute.ai",
    "ANTHROPIC_AUTH_TOKEN": "<your key>",
    "ANTHROPIC_MODEL": "qwen/qwen3.8-27b",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "qwen/qwen3.8-27b",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "qwen/qwen3.8-27b",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "qwen/qwen3.8-27b",
    "CLAUDE_CODE_MAX_CONTEXT_TOKENS": "131072"
  }
}

ANTHROPIC_AUTH_TOKEN takes the bare key; Claude Code adds Bearer. The three ANTHROPIC_DEFAULT_* variables send Claude Code's background and sub-agent requests to the same model, because a claude-* model id is refused. CLAUDE_CODE_MAX_CONTEXT_TOKENS gives Claude Code the real window. Without it, Claude Code assumes 200,000 tokens and compacts too late.

Run Claude Code in your project: claude. To run one prompt and exit: claude -p "Summarise this repository".

Thinking

Claude Code asks for thinking on its main requests, so the model thinks before it answers, even on a short question. That takes longer, and thinking tokens are billed as output.

To turn thinking off, add export CLAUDE_CODE_DISABLE_THINKING=1, or the same key in the env block. Claude Code then sends no thinking setting, and the model answers directly. Remove the variable to turn thinking back on.

Muse Glimmer 30B always reasons before it answers. The thinking that Claude Code asks for on its main requests gives Glimmer's own default, high; with CLAUDE_CODE_DISABLE_THINKING=1 it reasons briefly.

What does not work

Web search. Claude Code's WebSearch is a tool that Anthropic runs, so a request for it is refused. WebFetch works: Claude Code fetches the page itself.
PDFs, and images on the Qwen models. The Qwen models are listed for text only. A PDF reaches the model as a note that the file was left out.
Thinking budgets. MAX_THINKING_TOKENS and thinking signatures have no effect. Thinking is on or off, and the reply is still held to its output limit.
Muse Glimmer 30B on background requests. Some of Claude Code's background requests allow only 64 output tokens, which its reasoning can use up before any answer. If you choose it, keep ANTHROPIC_DEFAULT_HAIKU_MODEL on qwen/qwen3.8-27b.
Cache controls. Claude Code's caching marks are accepted and ignored. Caching is automatic: a repeated prefix shows as cache reads and is billed at the cached price.
Claude Code's cost display. It prices a model it does not know at its own rates. Your bill uses the prices in GET /v1/models.
Remote Control, which Claude Code turns off when the base URL is not Anthropic's, and features that ask for a Claude model by name, as reported of the VS Code extension.
Help from Anthropic. Anthropic's docs say it does not support Claude Code with models other than Claude. For this setup, write to [email protected].

Troubleshooting

API Error: 401 invalid API key: The key is wrong or missing. ANTHROPIC_AUTH_TOKEN takes the bare key, without Bearer. A value in the env block of a settings file wins over the shell, so check both.
There's an issue with the selected model: A model id is misspelled: we answer 404 and Claude Code shows its own message. A 400 that names ANTHROPIC_MODEL means a variable still names a claude-* model. Set ANTHROPIC_MODEL and the three ANTHROPIC_DEFAULT_*_MODEL variables to qwen/qwen3.8-27b.
Prompt is too long: The conversation is longer than the 131072-token context window. Check that CLAUDE_CODE_MAX_CONTEXT_TOKENS=131072 is set, so Claude Code compacts in time, then run /compact or /clear.
402 or 429: A 402 means your prepaid credit does not cover the request: add credit at creeksidecompute.ai/account. A 429 means the key's rate, or the number of requests your account may run at once, is used up; agents that work in parallel reach that sooner. Wait and retry. The API reference lists the limits.

Other tools

Sources, read on 2026-10-11: environment variables, LLM gateway, model configuration.