TopolearnMany apps and libraries only speak the OpenAI API; compatibility lets them use local models unchanged.
Many tools and libraries were written for OpenAI's API. Ollama also offers an OpenAI-compatible API under http://localhost:11434/v1/, so you can point that code at Ollama by changing two settings:
base_url(the server address) tohttp://localhost:11434/v1/,api_keyto any non-empty string, such as"ollama". It is required by the client library but ignored by a local server.
Supported endpoints include:
/v1/chat/completions: chat, streaming, JSON mode and structured output, tools, and base64 images./v1/completions: plain text completions./v1/embeddings: embeddings./v1/models: list models./v1/responses: OpenAI's newer Responses API, without its stateful features (you send the full history yourself).
Limitations to know: some parameters aren't supported (for example logprobs, tool_choice, n), and images must be sent as base64 data rather than URLs. There is also no parameter for context length. To use a bigger context through this API, create a custom model with PARAMETER num_ctx in a Modelfile, or set OLLAMA_CONTEXT_LENGTH on the server.
Ollama also has an Anthropic-compatible endpoint for tools that use Anthropic's Messages API, and the same /v1 paths work against https://ollama.com/v1 for cloud models with an API key.
Native /api/... endpoints expose more Ollama-specific options (keep_alive, think, every model option), so prefer them, or the ollama library, when you don't need OpenAI compatibility.
Example
# pip install openai
from openai import OpenAI
client = OpenAI(base_url="http://localhost:11434/v1/", api_key="ollama") # key is ignored locally
resp = client.chat.completions.create(
model="gpt-oss:20b",
messages=[{"role": "user", "content": "Say this is a test"}],
)
print(resp.choices[0].message.content) # This is a test.
emb = client.embeddings.create(model="embeddinggemma", input=["hello", "world"])
print(len(emb.data), len(emb.data[0].embedding)) # 2 768
print([m.id for m in client.models.list().data]) # ['gpt-oss:20b', 'embeddinggemma:latest', ...]Remember
- Use
base_url="http://localhost:11434/v1/"and a dummyapi_key - Chat, completions, embeddings, models and (stateless) Responses are supported
- No context-length parameter: use a Modelfile or OLLAMA_CONTEXT_LENGTH
Watch out for
- Expecting num_ctx in an OpenAI-compatible request to change the context; set it via a Modelfile or the server instead.