Ollama is useful from a terminal, but many apps expect an OpenAI-style HTTP API. Ollama provides one. If the app can point at a different base URL, you can often send that traffic to a model running on your own PC instead of to a cloud account.
Forge PC has not measured tokens per second through this API, and we have not audited every third-party client. The endpoints, fields and examples below come from Ollama's official OpenAI compatibility documentation.
This page assumes Ollama is already installed and running on Windows. If it is not, use our Ollama and LM Studio install guide first. For choosing how much memory the model needs, see how much RAM local AI uses and how much VRAM you need.
Start the local server
On a normal desktop install, Ollama's local server listens on http://localhost:11434. The OpenAI-compatible routes sit under /v1/, for example:
http://localhost:11434/v1/chat/completions
Official examples set an API key in the client because many OpenAI libraries require the field. Ollama ignores that key for local use; the docs use the placeholder value ollama.
Download a model before you call it. The OpenAI-compatibility page shows:
ollama pull llama3.2
Use whatever model you actually have. The model name in the JSON must match a name from ollama list. If you need a name that a tool insists on, such as gpt-3.5-turbo, Ollama documents ollama cp llama3.2 gpt-3.5-turbo as a way to copy an existing local model to that label. That does not download an OpenAI model; it only renames a local one.
A first request with curl
Ollama's documented curl example (with llama3.2 as the model name once you have pulled it) looks like this:
curl http://localhost:11434/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "llama3.2",
"messages": [
{
"role": "user",
"content": "Hello!"
}
]
}'
If Ollama is not running, or the firewall blocks port 11434, the request will fail. A simple check from the official docs is to confirm the server responds on that port.
Python and JavaScript clients
The official Python example creates an OpenAI client pointed at the local server:
from openai import OpenAI
client = OpenAI(
base_url='http://localhost:11434/v1/',
api_key='ollama', # required but ignored
)
chat_completion = client.chat.completions.create(
messages=[{'role': 'user', 'content': 'Say this is a test'}],
model='llama3.2',
)
print(chat_completion.choices[0].message.content)
The JavaScript example is the same idea: set baseURL to http://localhost:11434/v1/ and apiKey to ollama.
You can also use /v1/responses with client.responses.create(...) and an input string. Ollama added the Responses API in version 0.13.3 and documents it as the non-stateful flavour only: no previous_response_id or conversation support.
What the compatibility layer supports
According to the official compatibility page, local /v1/chat/completions supports chat, streaming, JSON mode, reproducible outputs, vision (base64 images, not remote image URLs), tools, and reasoning controls for thinking models. It does not support logprobs, tool_choice, logit_bias, user or n.
Other documented endpoints include /v1/completions, /v1/models, /v1/embeddings, /v1/audio/speech (from Ollama v0.40.8) and /v1/audio/transcriptions. The page lists which request fields each path accepts. Do not assume every OpenAI feature exists; if a field is marked unsupported, treat a 400 error as expected.
The OpenAI API has no field for context length. If you need a different context size, Ollama says to create a Modelfile with PARAMETER num_ctx and ollama create a new model name, then call that name.
Keep this server on your own PC unless you have deliberately followed Ollama's guidance for authentication and remote access. The basic local examples are for localhost, not a public port.
Ollama also documents a cloud API at https://ollama.com/v1 that uses OLLAMA_API_KEY. That is a different product from the local server this guide is about.
Audio, embeddings and listing models
If you want to see which models the server currently exposes in OpenAI form, call GET /v1/models. Ollama notes that created is the time the model was last modified, and owned_by is the Ollama username, defaulting to "library".
/v1/embeddings accepts a string or an array of strings, plus encoding format and dimensions. It does not accept arrays of tokens.
From Ollama v0.40.8, /v1/audio/speech can generate audio with speech models. The documented defaults matter: audio is 16-bit mono at 24kHz, the default response_format is wav rather than mp3, speed is 1 only, and unsupported values return a 400 error. From v0.40.12, /v1/audio/translations can translate speech into another language, with to_language defaulting to English.
/v1/audio/transcriptions accepts a WAV file (up to 384kHz, 512MB and two hours, per the docs), a model name, an optional language code, and json, text or verbose_json as the response format. Audio longer than the model takes at once is split at pauses. These paths only work with models that actually accept audio; a plain chat model will not suddenly become a transcription engine.
Things that usually go wrong
Connection refused. Ollama is not running, or something else is bound to port 11434. Start the app or service and try again.
404 model not found. The name in the JSON does not match ollama list. Pull the model, or fix the spelling and tag.
The client demands an API key. Supply a dummy value. Local Ollama ignores it.
A feature silently missing, or a 400 error. Check the official supported-fields list. tool_choice, logprobs and stateful Responses (previous_response_id) are not supported.
Context too short. You cannot set num_ctx in the OpenAI request. Create a derived model with a Modelfile, as the official page describes.
Do not paste API examples that talk to https://ollama.com/v1 into a local script and expect them to hit your PC. That base URL is the cloud service and needs OLLAMA_API_KEY.
Our verdict
Point your OpenAI-compatible client at http://localhost:11434/v1/, put any placeholder in the API key field, pull a model, and call /v1/chat/completions with that exact model name. Use the official compatibility list to see which fields work, and do not expose the port to the internet as an afterthought. We have not benchmarked clients through this API; the steps come from Ollama's official documentation.
