OpenAI Agents SDK - Hostlinku Docs

The OpenAI Agents SDK uses the Responses API by default, which is provided by Hostlinku at POST /v1/responses. Simply provide the OpenAI client with Hostlinku’s base_url, and disable tracing (tracing is exported to the OpenAI platform, not through the gateway).

How the agent loop works

Each iteration of the agent loop is a single request to the Hostlinku Gateway: the agent sends a prompt, the model responds with a tool call, the agent runs the tool locally (reads/writes files, executes commands), and then sends the results back—and so on until the model provides the final answer. The gateway routes each iteration to the upstream model, and everything is logged separately in the console → Request Log, complete with the token and associated cost.

Hostlinku agent loop diagram: IDE/coding agent, Hostlinku Gateway, upstream model, Hostlinku Console, and local workspace

Without changing the code: environment variable

The OpenAI SDK reads OPENAI_BASE_URL and OPENAI_API_KEY, so the Agents SDK can be directed to Hostlinku without modifying the code—just disable tracing via the environment variables as well:

export OPENAI_BASE_URL=https://api.hostlinku.com/v1
export OPENAI_API_KEY=sk-isi-api-key-kamu          # API key Hostlinku
export OPENAI_AGENTS_DISABLE_TRACING=1
python agent.py

Responses API or Chat Completions?

The default SDK is the Responses API (/v1/responses) — recommended, as it supports reasoning items and offers the most comprehensive set of tool workflows. If you want Chat Completions (/v1/chat/completions):

from agents import set_default_openai_api
set_default_openai_api("chat_completions")

import { setOpenAIAPI } from "@openai/agents";
setOpenAIAPI("chat_completions");

Both use the same API key, balance, and request logs.

Supported via the gateway

Agents SDK Features	Status
`function_tool` / `tool()`, parallel tool calls	Yes
Handoffs between agents (including different models per agent)	Yes — mix `claude-sonnet-4-6` and `gpt-5-5` in a single workflow
Structured output (`output_type` / Zod)	Yes, for models that support JSON schema
Streaming (`Runner.run_streamed`, `run(..., { stream: true })`)	Yes
`reasoning={"effort": …}` via `ModelSettings`	Yes, passed to the model
OpenAI hosted tools (`WebSearchTool`, `FileSearchTool`, `CodeInterpreterTool`)	Depends on the target model; run the tool as its own `function_tool` if it is not available
Tracing to the OpenAI platform	Not via the gateway — disable it, or use a third-party trace processor

Troubleshooting

Full error log tracing export 401 — tracing has not been disabled; call set_tracing_disabled(True) / setTracingDisabled(true) or set OPENAI_AGENTS_DISABLE_TRACING=1.
400 previous_response_id is not supported — remove the use of previous_response_id / conversation_id; see the callout above.
404 model_not_found — model= must be the Hostlinku slug, not the original OpenAI/Anthropic model name.

​How the agent loop works

​Without changing the code: environment variable

​Responses API or Chat Completions?

​Supported via the gateway

​Troubleshooting

How the agent loop works

Without changing the code: environment variable

Responses API or Chat Completions?

Supported via the gateway

Troubleshooting