Keep your OpenAI code. Change two values.

The two values.

Migration from an existing OpenAI integration is a configuration change, not a rewrite. Replace the endpoint host and the credential, keep every request shape and retry helper you already tested. The catalog, billing formula and request log stay the same for all models.

Python uses the first console block on this page. The same two values work for curl when the header carries the bearer key and the path targets /v1/chat/completions. Keep timeouts, retries and logging from your current client; only the host and credential change.

curl https://serveras.click/v1/chat/completions -H "Authorization: Bearer asai_sk_..." -H "Content-Type: application/json" -d '{"model":"asai/gpt-6-sol","messages":[{"role":"user","content":"Hello"}]}'

Teams on Codespaces, shared runners or office machines in Kuala Lumpur and beyond can store the key once as a repository secret and reuse this exact block across services. Nothing in the request identifies a region; billing stays in MYR no matter where the code executes.

List models live.

Model aliases retire and replacements arrive without a version bump, so hardcoding a list invites 404 errors weeks later. Call GET /v1/models at deploy time or in a startup check, then select an id from the response. Use the value verbatim, prefix included, because gpt-6-sol without the asai/ prefix is not the same alias.

The catalog response also carries context length, maximum output, input modalities and billing mode per alias. Build prompts against those limits instead of assuming every model accepts images or long documents. The browser view at Models renders the same document in a searchable table.

Document input follows the same schema: an alias whose input_modalities includes file accepts a {"type": "file"} content part carrying a base64 data URL. One limit applies across the catalog: 100,000 characters of document text per request, at most five documents. An oversized document is rejected with a 400 before any credit is reserved, so nothing is charged for it.

Consistent request handling.

Because the gateway faithfully implements the OpenAI Chat Completions schema, features like streaming, structured JSON output and tool calling work as expected when the underlying model supports them. The response structure remains identical whether you use Kimi K3, DeepSeek V4 or a local free model, ensuring your parsing logic never needs conditional branching based on the provider.

When an error occurs, the gateway translates provider-specific failures into standard HTTP status codes and OpenAI-style error objects. A depleted balance returns a standard 402 Payment Required, a rate limit hit returns a 429, and a missing model returns a 404, allowing your existing exponential backoff and retry handlers to function normally without new rules.

Claude Code and Codex note.

Hermes, Claude Code and Codex all speak to asAi through the same key and prepaid balance, but each client expects a different compatibility document. Claude Code follows the Messages-compatible flow in the Claude Code guide, while Codex follows the Responses-compatible flow in the Codex guide. Use this page for plain OpenAI SDK traffic and switch to the dedicated guide when the client complains about request shape.

Malaysian developers commonly keep one provider entry per client rather than one shared entry. That habit makes per-client failures easier to isolate in Logs, because each request carries its own model alias and timestamp.

Secure your key.

Treat the key like a database password. Read it from the environment at process start, scope repository secrets to the environments that need them, and never paste it into chat, screenshots, tickets or generated code samples. A key that appeared anywhere public should be revoked in the dashboard and replaced immediately.

In GitHub Codespaces, store the value as a codespace secret and export it into the shell before running the SDK, instead of writing it into a checked-in .env file. Each developer can hold a labelled key under API keys, and the account stays under the three-key limit when old keys are revoked after rotation.

Billing in credits.

The same prepaid balance funds every compatible client, so a Hermes session, a Claude Code run and a direct SDK call all deduct from one pool. Read the multiplier table in Billing, estimate with the examples above, and confirm actual spend in Logs before scaling traffic.

OpenAI-compatible API FAQ.

Paying in ringgit instead? See AI API for Malaysia. Connecting an agent? Read Setup Hermes Agent di Malaysia dengan asAi or the English technical guide.