Local models

aeoscope can query models running on your own machine - no API keys, no data leaving your network. Two ways to connect:

Ollama (zero configuration)

1. Install Ollama and pull a model:

   ollama pull llama3.2

2. Enable the Ollama engine in aeoscope's Settings. That is it - aeoscope talks to Ollama's default address (http://127.0.0.1:11434).

To use a different model, set it in your .env:

OLLAMA_MODEL=qwen2.5

LM Studio, llama.cpp, vLLM, and everything else

Almost every local inference server exposes an OpenAI-compatible API. The Custom engine speaks that format, so one setting covers all of them.

LM Studio: load a model, then start the server from the Developer tab (default port 1234):

CUSTOM_BASE_URL=http://localhost:1234/v1

llama.cpp server:

CUSTOM_BASE_URL=http://localhost:8080/v1

vLLM:

CUSTOM_BASE_URL=http://localhost:8000/v1

Ollama also has an OpenAI-compatible endpoint, if you prefer to use it through the custom engine:

CUSTOM_BASE_URL=http://localhost:11434/v1

Then set the variables in your .env, restart aeoscope, and enable the Custom engine in Settings:

CUSTOM_BASE_URL=http://localhost:1234/v1
CUSTOM_MODEL=google/gemma-4-e4b  # the model's identifier in your server (see below)
CUSTOM_API_KEY=                  # only if your server requires a key

Set CUSTOM_MODEL to the exact identifier of a loaded model - in LM Studio you can find it in the Developer tab's server view or via lms ls (e.g. google/gemma-4-e4b, qwen/qwen3.5-9b). If unset, aeoscope sends local-model, which servers that always answer with their loaded model accept, but LM Studio can reject when it does not match.

Running aeoscope in Docker

When aeoscope runs in Docker (docker compose up) and your model server runs directly on your machine, localhost inside the container refers to the container itself - not your machine. Point the base URL at the host using Docker's special host.docker.internal name:

Server	On your machine	From aeoscope in Docker
LM Studio	http://localhost:1234/v1	http://host.docker.internal:1234/v1
llama.cpp server	http://localhost:8080/v1	http://host.docker.internal:8080/v1
vLLM	http://localhost:8000/v1	http://host.docker.internal:8000/v1
Ollama	http://localhost:11434 (Ollama engine) or :11434/v1 (Custom engine)	http://host.docker.internal:11434 or :11434/v1

Two things make this painless:

• A localhost URL still works. When a request to a loopback address fails, aeoscope automatically retries once via host.docker.internal and logs a warning with the URL to set. Prefer the correct URL so the failed attempt is skipped.
• Linux needs one extra toggle. On Docker Desktop (Windows/macOS) everything above works as-is. On Linux, the provided docker-compose.yml maps host.docker.internal to the host, but a server bound to 127.0.0.1 only listens on loopback - enable its "serve on local network" option (bind to 0.0.0.0) so Docker can reach it.

After changing .env, recreate the container with docker compose up -d - the container only reads environment variables when it is created.

What local models can and cannot tell you

• Hosted engines with web access (Perplexity especially) show what real users actually see when they ask about your category. That is the core AEO signal.
• Local models have no web access and fixed training data. They give you offline, zero-cost, unlimited-rate signals - useful for testing pipelines, comparing prompt phrasings, and keeping scheduled runs running cheaply between hosted checks.

Use both: local engines for volume, hosted engines for truth.