One API. Your language.

Install Claude Code from Claude's setup guide, set your key as above, then start it from that terminal. Keep keys out of project files.

This example uses Sonnet 4.6. It turns off adaptive thinking and experimental betas.

(
  export ANTHROPIC_BASE_URL=https://api.tokenspend.dev
  export ANTHROPIC_AUTH_TOKEN="$TOKENSPEND_API_KEY"
  unset ANTHROPIC_API_KEY
  export ANTHROPIC_MODEL="anthropic/claude-sonnet-4-6"
  export CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1
  export CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1
  claude
)

For a non-Claude model, also set MAX_THINKING_TOKENS=0. For a smaller output limit, such as Nova's 5,000, set CLAUDE_CODE_MAX_OUTPUT_TOKENS to it.

In /status, check the base URL and the credential source. A settings file can override both. Exit the subshell to go back to your normal settings.

Adaptive thinking, server-side tools and unlisted beta fields return a 400. Keep tool results in Messages format.

Sources: gateway authentication · feature compatibility.

Codex talks to custom providers over the Responses API. Add TokenSpend as a provider in your Codex config, with the key in TOKENSPEND_API_KEY.

model_provider = "tokenspend"

[model_providers.tokenspend]
name = "TokenSpend"
base_url = "https://api.tokenspend.dev/v1"
env_key = "TOKENSPEND_API_KEY"
wire_api = "responses"

Function tools, tool-result continuation and reasoning controls work on Astra and the GPT-5 models. Stored-response chaining and hosted tools are not supported.

See Codex provider configuration and our Responses support.

1. Open Cursor Settings → Models. If your version has Override OpenAI Base URL, set it to https://api.tokenspend.dev/v1. If it does not, this setup is not available in your version.
2. Paste a TokenSpend key into the API-key field, never a Cursor admin key. Add an exact TokenSpend model ID where custom models are allowed.
3. Send a small chat request and check its receipt before you use Agent mode. Restore the old URL and key when you switch back.

Tab completion does not use custom keys. Cursor sends API-key requests through its own servers, so TokenSpend logging settings do not cover that path.

Sources: Cursor API keys.

Run opencode --version, then merge the matching config into your project's OpenCode config without overwriting your settings. Start OpenCode from a terminal with your key set.

{
  "$schema": "https://opencode.ai/config.json",
  "model": "tokenspend/zai/glm-5.2",
  "provider": {
    "tokenspend": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "TokenSpend",
      "options": {
        "baseURL": "https://api.tokenspend.dev/v1",
        "apiKey": "{env:TOKENSPEND_API_KEY}"
      },
      "models": {
        "zai/glm-5.2": {
          "name": "zai/glm-5.2"
        }
      }
    }
  }
}

{
  "$schema": "https://opencode.ai/config.json",
  "model": "tokenspend/coder",
  "providers": {
    "tokenspend": {
      "env": [
        "TOKENSPEND_API_KEY"
      ],
      "package": "@opencode/ai/providers/openai-compatible",
      "settings": {
        "baseURL": "https://api.tokenspend.dev/v1"
      },
      "models": {
        "coder": {
          "modelID": "zai/glm-5.2"
        }
      }
    }
  }
}

Pick the TokenSpend entry in /models. It uses the Chat Completions API. Tool results must reference the matching call IDs, and unsupported fields return a clear error.

Sources: V1 custom provider · V2 provider configuration.

1. Send one small request with synthetic content. It is billable. Note the request ID, not the key.
2. Open Router activity in the same workspace and match the request ID, model, status, token counts and cost.
3. With an agent, ask it to read a harmless README, summarize it, then answer a follow-up. You should see a tool call, a tool result and matching receipts.

Nothing showing up? Check the error code and API status. Contact us with the request ID, UTC time, model and client version. Do not send keys or private prompts.