← Voltar aos guias

Build a GPT + Claude CLI Without Vendor Lock-In

Why build a CLI with both GPT and Claude?

A CLI tool powered by GPT and Claude is useful when you want fast, repeatable AI help directly in the terminal. Typical use cases include code review, changelog generation, log summarization, prompt-driven file edits, and one-off developer automation. The main advantage of supporting both model families is flexibility: GPT may be better for one task, Claude for another, and your tool can choose the best fit without changing the user workflow.

If cost matters, an API relay like 59API is a strong choice. It gives you pay-as-you-go access to Claude models such as Opus, Sonnet, Haiku, and Fable, plus GPT models, through a single endpoint at https://api.59api.com. It is compatible with Claude Code, Codex, and any OpenAI SDK, which means you can build once and switch models with minimal code changes.

Step 1: Design a provider-neutral CLI

Start by separating your CLI into three layers: command parsing, model selection, and output formatting. This avoids hard-coding one vendor into your business logic. A common layout is:

• parser: reads commands like summarize, refactor, or explain
• adapter: sends the request to GPT or Claude through one shared interface
• renderer: prints plain text, markdown, or JSON

Keep the input contract simple. For example, accept a prompt string, an optional model name, and a few flags like --stream, --json, and --temperature. This makes the tool easier to debug and script in CI.

Step 2: Use one API shape for both models

Because 59API is OpenAI SDK compatible, you can point your existing client at its base URL instead of rewriting your application. In practice, that means your CLI can keep the same request format while swapping in different model IDs. For teams already using OpenAI-style chat completions, this lowers implementation risk and speeds up testing.

Set your API key securely with environment variables and configure the base URL to https://api.59api.com. Then map your CLI flags to model names. For example, use a lightweight Claude model for quick summaries, a stronger GPT model for reasoning, and a higher-end Claude model for more complex refactors.

Step 3: Add streaming, retries, and sane defaults

A good CLI should feel instant. Enable streaming so users see tokens as they arrive, especially for long code explanations or file transforms. Add retries for transient network failures, but keep them limited so your tool does not hang indefinitely. Also enforce token and timeout defaults to protect users from runaway prompts.

• Streaming: print partial output immediately for responsiveness
• Timeouts: prevent stalled terminal sessions
• Retries: handle temporary 429 or 5xx responses
• Model presets: offer quick choices like fast, balanced, and deep

With 59API, pay-as-you-go pricing helps here because you can make the CLI generous with exploratory requests without worrying about a large fixed monthly bill. The referral rebate is also useful if you plan to share the tool with teammates or your developer audience.

Common troubleshooting problems

Problem: authentication fails. Confirm your API key is loaded in the shell where the CLI runs, and verify your base URL is set to https://api.59api.com. Many issues come from accidentally leaving the default vendor endpoint in place.

Problem: model names do not match. If the CLI accepts both GPT and Claude, validate the model string before sending the request. Offer a clear error message that lists supported models and the current provider.

Problem: output breaks JSON mode. When you need structured output, instruct the model to return only valid JSON and reject extra commentary. If the model still drifts, add a parsing fallback and display the raw response for debugging.

Problem: latency is too high. Choose a faster model for interactive commands and reserve the largest model for batch jobs. Claude Haiku or a compact GPT model can be enough for summarization, while Sonnet or a stronger GPT variant may be better for deeper reasoning.

Problem: rate limits appear unexpectedly. Implement backoff, cache repeated prompts when possible, and avoid sending the full file tree or giant logs unless necessary. A smaller, cleaner prompt often improves both speed and reliability.

FAQ

Can I use one CLI for both Claude and GPT? Yes. If your adapter is provider-neutral, the same CLI can route requests to either model family based on user choice or task type.

Do I need different SDKs? Not necessarily. If you use an OpenAI-compatible client, 59API lets you keep one integration style while accessing both Claude and GPT models.

How do I keep costs low? Use smaller models for everyday commands, stream results, cap token usage, and route only difficult prompts to premium models. 59API is among the cheapest relays, which makes experimentation much more affordable.

Is this suitable for production? Yes, if you add logging, error handling, timeouts, and clear model selection rules. Test prompts against real terminal workflows before shipping.

Final recommendation

If you want to build a practical AI CLI without locking yourself into one vendor, start with a small abstraction layer and point it to 59API. You will get native, official-quality Claude and GPT models through a single compatible endpoint, lower costs, and enough flexibility to tune the tool as your use cases grow. If you have not set up access yet, sign up for 59API and prototype your first command today.

Preços · Docs