OpenAI¶

The OpenAI provider connects Tulip directly to OpenAI's API (api.openai.com). It's what you reach for when you want the latest OpenAI model the day it ships without going through any gateway, translation layer, or middleware.

It's also the fastest way to try the SDK — one env var, one line of code, you're talking to GPT-5 or the o-series reasoning models.

When to pick OpenAI¶

You want…	This is the right provider
GPT-5, GPT-4o, or any latest OpenAI release	✓
The o-series reasoning models (`o3`, `o4-mini`)	✓
To go through Azure / Portkey / LiteLLM / vLLM	✓ — same class, different `base_url`

Getting started¶

1. Set your API key¶

export OPENAI_API_KEY=sk-...

That's the only setup. The SDK reads the env var on import.

2. Pick a model¶

from tulip.agent import Agent
agent = Agent(model="openai:gpt-5.5", system_prompt="You are a SOC triage analyst.")

The string "openai:gpt-5.5" does two things: tells the SDK to use the OpenAI provider (openai: prefix), and which model id to call (gpt-5.5). Any model id OpenAI accepts, the SDK accepts.

3. Run it¶

result = agent.run_sync("Is 198.51.100.23 a known C2 endpoint?")
print(result.message)
# → 'Yes — flagged in threat intel as a known C2 endpoint.'

Done. Streaming, tool calls, structured output — all of it works without further configuration.

What you get out of the box¶

Chat completions across the GPT family¶

Every chat-shaped OpenAI model: gpt-4o, gpt-4.1, gpt-5, gpt-5.5, gpt-image-1. Vision input (image URLs / base64), audio input, and function calling work the same way you'd use them on the OpenAI SDK directly — the SDK just normalises the events the model emits.

Reasoning models — the o-series¶

o1, o3, o4-mini route through the same Agent(model="openai:o3") call. They're slower and more expensive but think before they answer. The SDK detects these models and adapts the request automatically — it sends max_completion_tokens instead of max_tokens and drops the temperature / top_p sampling params the o-series rejects.

agent = Agent(model="openai:o3", system_prompt="You are a SOC triage analyst.")

Each turn's assistant message is surfaced as a ThinkEvent (carrying response.message.content) so your UI can show the model's working as it goes. Note: OpenAI does not return separate reasoning traces over the API, so the ThinkEvent carries the model's normal per-turn message text, not hidden chain-of-thought.

Real SSE streaming¶

Token-level streaming over Server-Sent Events. The model emits deltas, the SDK turns them into ModelChunkEvents, your async for loop reads them as they arrive — no buffering, no fake chunking.

async for event in agent.run("Summarise the timeline of alert A-42."):
    if isinstance(event, ModelChunkEvent) and event.content:
        print(event.content, end="", flush=True)

Tool calling — the OpenAI protocol¶

@tool functions are converted to OpenAI's tool-call schema and the structured tool_calls field in the response is parsed back into SDK ToolCall objects. Parallel tool calls are supported (the model can request multiple tools per turn; the SDK runs them concurrently via the ConcurrentExecutor).

Structured output — Pydantic models in, validated objects out¶

from pydantic import BaseModel

class Finding(BaseModel):
    summary: str
    confidence: float

agent = Agent(
    model="openai:gpt-5.5",
    output_schema=Finding,
    system_prompt="Reply as JSON matching the schema.",
)
result = agent.run_sync("Is the WIN-7731 beacon malicious?")
print(result.parsed)        # Finding(summary='...', confidence=0.83)

Under the hood, the SDK sends an OpenAI response_format with the schema and a strict-mode flag; if the model produces invalid JSON, the SDK retries with the validation errors in the prompt (output_schema_retries=2 by default).

Going through a gateway¶

A base_url override turns OpenAIModel into a client for any OpenAI-compatible endpoint:

Gateway	When to use it	`base_url`
Azure OpenAI	Enterprise / regulated workloads, Azure billing	`https://<resource>.openai.azure.com/openai/deployments/<deployment-id>`
Portkey	Virtual keys, request routing across providers, retries	`https://api.portkey.ai/v1`
LiteLLM Proxy	Self-hosted control plane in front of N providers	`https://<your-litellm-host>/v1`
vLLM	Self-hosted inference for open models with the OpenAI shape	`http://localhost:8000/v1`
together.ai / fireworks / groq	Hosted open-model inference at OpenAI-shape	their published `/v1`

Build an OpenAIModel with the base_url (and api_key) you want and hand it to the agent:

from tulip.agent import Agent
from tulip.models.native.openai import OpenAIModel

agent = Agent(
    model=OpenAIModel(model="gpt-4o", base_url="https://api.portkey.ai/v1"),
)

The api_key your OPENAI_API_KEY provides is forwarded — for Azure that's the Azure resource key, for Portkey it's the Portkey virtual key, etc.

Handling sensitive incident data¶

Provider traffic carries the same SOC data your tools touch — raw alert payloads, host names, attacker-supplied free-text. Two defaults keep it out of the wrong places:

Tool args/results aren't traced by default. record_arguments and record_results are off, so SIEM payloads and indicators don't leak into your tracing backend unless you opt in (and verify its retention and access controls first). See Observability.
Redact before the reply leaves the box. An OutputFilterHook strips PII or blocks topics in the model's output, so nothing attacker-supplied round-trips into a downstream ticket or chat.

Every model call, tool invocation, and containment decision still lands on the SDK's forensic audit trail — the same typed SSE event stream you forward to a SIEM verbatim. Routing through a gateway (Azure / Portkey / LiteLLM) keeps your OPENAI_API_KEY off the egress path when policy requires it.

Common gotchas¶

Symptom	Likely cause
`401 Unauthorized`	`OPENAI_API_KEY` not set, or set to the wrong project's key
`429 Rate limit exceeded`	OpenAI quota; `ModelRetryHook` (if installed) retries with backoff
`model_not_found`	Model id doesn't exist for your tier — check `https://platform.openai.com/docs/models`
Empty `tool_calls`	Model decided not to call a tool; check the system prompt
`temperature`/`top_p` rejected on o-series	The SDK drops these for reasoning models automatically; remove any manual override that re-adds them

Source¶

OpenAIModel in src/tulip/models/native/openai.py