Claude Code

Run Claude Code on Unspent credits with two environment variables.

Claude Code runs on Unspent credits through its gateway settings. Set the variables and start it:

export ANTHROPIC_BASE_URL=https://api.egrtgghfghtytgb.space
export ANTHROPIC_AUTH_TOKEN=usp_live_...
export CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1
claude

The third line is optional: it lists Unspent's models in /model.

To keep the setup across sessions, put the same values in ~/.claude/settings.json:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.egrtgghfghtytgb.space",
    "ANTHROPIC_AUTH_TOKEN": "usp_live_..."
  }
}

Choose a model with /model or claude --model claude-opus-5-5. The list: Models & pricing.

What to expect

• Each request waits for its hold to confirm on Solana before the first token. That adds the confirmation time of one transaction to every request.
• Prompt caching works as usual. Claude Code reads most of its context from the cache, at the cache read price.
• Tools that Claude Code runs on your machine, such as bash and file edits, are billed as normal tokens.
• Server tools Unspent doesn't offer, such as web fetch, are rejected in the provider's wording, and Claude Code carries on without them.
• Fast mode is not available.
• /app/usage shows spend per Claude Code session.

Interrupting a request

Pressing Esc stops Claude Code from reading the answer, but the model keeps generating it. Unspent reads the stream to its end and bills everything the model generated, up to max_tokens. An interrupted request costs as much as one you let finish.

Many requests at once

Subagents send requests in parallel. A wallet can have 50 credits held at once, or less if you set a lower hold limit. Requests over it get 429 too_many_in_flight with retry-after, and Claude Code retries them. See Rate limits.

Running out of credits

When a request's hold doesn't fit your available balance, Claude Code shows Not enough credits. Available: …. Buy more on /app/buy and continue the session.