← Back to all guides

Point Any AI CLI at a Custom Base URL Safely

Why Custom Base URLs Often Fail

Pointing an AI CLI at a custom API endpoint sounds simple: change one environment variable, add an API key, and run a prompt. In practice, small configuration mismatches can cause confusing 401, 404, model-not-found, or unsupported-endpoint errors. The main issue is that AI command-line tools do not all use the same provider protocol, variable names, or URL formatting rules.

A reliable low-cost endpoint is useful when you want to keep your existing workflow while reducing model costs. 59API is an AI API relay with pay-as-you-go access to native official-quality Claude models, including Opus, Sonnet, Haiku, and Fable, plus GPT models. Its API base URL is https://api.59api.com, and it is designed for Claude Code, Codex, and OpenAI SDK-compatible workflows without a model downgrade.

Pitfall 1: Using the Wrong Environment Variable

The most common mistake is setting a variable that your CLI never reads. OpenAI-compatible tools commonly use OPENAI_BASE_URL and OPENAI_API_KEY. Anthropic-compatible tools may instead expect ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN. Some CLIs use a config file, a named provider profile, or command-line flags rather than environment variables.

Before changing anything, identify the CLI's provider mode. If it is configured for an OpenAI-compatible API, set the OpenAI-style base URL and API key fields. If it speaks the Anthropic protocol, use its Anthropic-compatible settings. For example, a shell session might export OPENAI_BASE_URL=https://api.59api.com and then export OPENAI_API_KEY=your_59api_key. Do not assume that setting both OpenAI and Anthropic variables will make every tool work; it can make troubleshooting harder if the CLI selects an unexpected provider.

• Check the CLI's configuration command or documentation for the exact variable names.
• Start a new terminal session after editing shell profile files.
• Confirm the active value with your shell's environment inspection command before testing.

Pitfall 2: Adding or Removing URL Paths by Guesswork

Developers often append /v1, /api, or a trailing slash because another provider required it. That can produce a valid-looking URL that resolves to the wrong path. Use the documented 59API base URL, https://api.59api.com, exactly where your client asks for its base URL. Whether a CLI appends versioned paths itself depends on that CLI and its API adapter.

Pay attention to the difference between a base URL field and a full endpoint field. A base URL should normally contain only the relay origin supplied by the provider. A full endpoint field may require the complete request path. If a request returns 404, inspect the final URL in verbose logs before changing keys or models. A 404 usually indicates a path or protocol issue, not a billing problem.

Pitfall 3: Mixing API Protocols and Model Names

A custom URL does not automatically convert an Anthropic request into an OpenAI request, or vice versa. Claude Code needs an Anthropic-compatible route and authentication setup. Codex and OpenAI SDK-based tools generally need an OpenAI-compatible route, although their configuration formats vary by version. Select the provider mode that matches the CLI's request format first, then point that mode at 59API.

Model aliases are another frequent source of errors. A CLI may have a default model name that is unavailable through the selected provider profile. Explicitly configure a model offered by your account and compatible with the chosen protocol. For coding work, choose the appropriate Claude or GPT model available through 59API rather than relying on a hard-coded default from another vendor. If the CLI says a model is unavailable, verify the exact model identifier in the relay dashboard or documentation instead of inventing a similar name.

Pitfall 4: Testing With a Large Agent Job First

Do not validate a new endpoint by launching a long autonomous coding task. A bad base URL can create repeated retries, wasted tokens, and unclear logs. First run a minimal prompt that returns a short answer. Then test tool calling, file access, streaming, and the larger agent workflow you actually use.

• Test authentication with one short text request.
• Enable verbose or debug logging to view the target host, response status, and request path.
• Confirm the selected model in the CLI output.
• Only then run a repository-wide coding or refactoring task.

Pitfall 5: Forgetting Security and Team Configuration

Never paste relay keys into source code, committed configuration files, screenshots, or shell history. Store keys in environment variables or your team's approved secret manager. For shared development setups, document the required provider mode, base URL, model name, and key variable so teammates do not accidentally send requests to a different endpoint.

Once configured correctly, 59API can let you keep familiar AI CLI tools while using inexpensive, pay-as-you-go access to high-quality native Claude and GPT models. Its referral rebate can also reduce effective costs for developers who share the service responsibly. If you want to test a custom endpoint with your current CLI, sign up for 59API, create an API key, and begin with a small verification request before moving production workflows over.