Claude Code with teamToken

Claude Code starts and works against our endpoint: we accept Anthropic Messages requests, streaming and tools included. What answers them is GPT-5.5, GPT-5.3-codex or Gemini. Claude models themselves are not in our catalog, so if that is what you need, this page will not help.

The model situation, before you spend time on setup

Our Claude upstream fell through and those models are not in the catalog. Everything else about Claude Code keeps working: the interface, file edits, tools, the conversation, streamed output. What changes is which model sits at the other end.

So this page is for one case: you like the tool and want it pointed somewhere else, either because access and payment have stopped working where you are or because you would rather not hold a subscription for something you use in bursts. If the requirement is specifically Opus, the honest answer is not today.

Setup

Three variables. Put them in ~/.claude/settings.json and they survive a terminal restart.

// ~/.claude/settings.json
{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.teamtoken.store",
    "ANTHROPIC_AUTH_TOKEN": "<your teamToken key>",
    "ANTHROPIC_MODEL": "gpt-5.6-sol"
  }
}

Or for the current shell session only:

export ANTHROPIC_BASE_URL="https://api.teamtoken.store"
export ANTHROPIC_AUTH_TOKEN="<key>"
export ANTHROPIC_MODEL="gpt-5.6-sol"

The third one is not optional. Without it Claude Code asks for its default Claude model, which we do not have, and you get an error naming that model.

Checking that it works

Faster than launching the tool: the same endpoint answers a plain curl in Messages format.

curl https://api.teamtoken.store/v1/messages \
  -H "x-api-key: $TEAMTOKEN_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "gpt-5.6-sol",
    "max_tokens": 64,
    "messages": [{"role": "user", "content": "ping"}]
  }'

The reply comes back Anthropic-shaped, with content blocks and usage. The anthropic-version header is required, exactly as on the original API.

Which model for which work

The model switches on the fly, through ANTHROPIC_MODEL or the tool's own --model flag. The full list comes from GET /v1/models on the same key.

Spend caps per key

An agent in a loop can spend more than you planned, which is an ordinary Tuesday rather than a hypothetical. Each key carries its own dollar ceiling: issue one for Claude Code, cap it at $30, and $30 is where it stops regardless of what the task does.

The balance is shared across the whole catalog while the cap sits on the individual key, which means one runaway session cannot reach the money you set aside for anything else. Revoking a key is instant and leaves the others alone.

When it does not work