Ollama — run AI locally with screenpipe

Use Ollama to run AI models locally with screenpipe. Free, private, no API keys needed.

Ollama lets you run AI models locally on your machine. screenpipe integrates natively with Ollama — no API keys, no cloud, completely private.

why use Ollama with screenpipe?

• 100% local - all AI processing happens on your machine
• no API costs - free to use, no subscription required
• privacy - your screen data never leaves your computer
• offline - works without internet connection
• choice of models - pick from dozens of open-source models
• no rate limits - use as much as you want

setup

1. Install Ollama

   download and install from ollama.com

   supported platforms:

   • macOS (Apple Silicon & Intel)
   • Linux
   • Windows (via WSL)
2. Pull a model

   download a model to use:

   ollama run ministral-3
   

   this downloads the model and starts Ollama.

3. Select Ollama in screenpipe

   • open the screenpipe app
   • click the AI preset selector (top of chat/timeline)
   • click Ollama
   • pick your model from the dropdown
4. Start chatting

   ask screenpipe: “what did I work on this morning?”

screenpipe automatically detects Ollama running on localhost:11434.

recommended models

choose a model based on your hardware and needs:

fast & lightweight

model	size	RAM needed	best for
ministral-3	~2 GB	8 GB	fast, general use, great starting point
gemma3:4b	~3 GB	8 GB	strong quality for size, good for summaries
qwen3:4b	~3 GB	8 GB	multilingual, good reasoning
phi4	~3 GB	8 GB	fast, great for code

balanced

model	size	RAM needed	best for
llama3.3:8b	~5 GB	16 GB	strong all-around performance
deepseek-r1:8b	~5 GB	16 GB	excellent reasoning
mistral:7b	~4 GB	12 GB	good quality, widely used

high quality

model	size	RAM needed	best for
llama3.3:70b	~40 GB	64 GB+	best quality, needs high-end hardware
deepseek-r1:70b	~40 GB	64 GB+	best reasoning, needs high-end hardware
qwen2.5:32b	~20 GB	32 GB+	excellent quality, still usable on consumer hardware

specialized

model	size	RAM needed	best for
codellama:13b	~7 GB	16 GB	code generation and review
llava	~5 GB	16 GB	vision + language (can analyze screenshots)
mistral-openorca	~4 GB	12 GB	instruction following

pulling models

download any model from ollama.com/library:

# basic usage
ollama pull <model-name>

# examples
ollama pull ministral-3
ollama pull llama3.3:8b
ollama pull deepseek-r1:8b

# specific quantization (smaller = faster, larger = better quality)
ollama pull llama3.3:8b-q4_0  # 4-bit quantization
ollama pull llama3.3:8b-q8_0  # 8-bit quantization

model quantization explained

quantization reduces model size and speeds up inference:

• q4_0 - smallest, fastest, lower quality
• q5_0 - balanced
• q8_0 - larger, slower, better quality
• default (no suffix) - recommended balanced version

for most users, the default version is best.

using Ollama with screenpipe

in chat

once configured, use Ollama in screenpipe’s AI chat:

> what did I work on this morning?

> summarize my meeting from 2pm today

> find that documentation I was reading about React

> which apps did I use most yesterday?

screenpipe queries your screen/audio history and sends context to Ollama running locally.

in pipes

pipes (automations) can also use Ollama:

1. go to pipes in screenpipe sidebar
2. select a pipe (e.g., “day recap”, “time tracking”)
3. in pipe settings, select your Ollama preset
4. enable the pipe

now your automations run completely local and private.

performance tips

choose the right model for your hardware:

8 GB RAM → ministral-3, gemma3:4b, phi4
16 GB RAM → llama3.3:8b, deepseek-r1:8b, mistral:7b
32 GB+ RAM → qwen2.5:32b, larger models

optimize for speed:

• use 4-bit quantized models (q4_0)
• close other GPU-heavy applications
• use smaller context windows (less screen history)

optimize for quality:

• use 8-bit or full precision models
• use larger models (8b, 13b, or more)
• give more context in queries

GPU acceleration:

• Ollama automatically uses GPU if available
• NVIDIA GPUs: works out of the box
• AMD GPUs: supported on Linux
• Apple Silicon: uses Metal acceleration

troubleshooting

“ollama not detected”

• ensure Ollama is running: ollama serve
• check it’s responding: curl http://localhost:11434/api/tags
• verify Ollama is installed: ollama --version

model not showing in dropdown?

• pull it first: ollama pull ministral-3
• refresh screenpipe’s model list
• you can also type the model name manually

slow responses?

• try a smaller model (ministral-3, phi4)
• reduce context window (query shorter time ranges)
• close other GPU-heavy apps
• ensure you have enough free RAM (model size + ~2 GB overhead)

out of memory errors?

• use a smaller model
• use a quantized version (q4_0)
• close other applications
• check available RAM: model size + 2 GB minimum

model responses cut off?

• increase max tokens in Ollama settings
• some models have built-in limits
• try a different model

Ollama server crashes?

• check RAM usage (likely out of memory)
• try a smaller model
• restart Ollama: pkill ollama && ollama serve

can’t connect to Ollama?

• verify port 11434 is not blocked
• check Ollama is listening: lsof -i :11434
• try restarting Ollama

still stuck? ask in our discord — get model recommendations and troubleshooting help.

comparing models

want to test which model works best for you?

1. pull multiple models: ollama pull ministral-3 gemma3:4b llama3.3:8b
2. try the same query with each model
3. compare speed, quality, and RAM usage
4. stick with the one that fits your needs

requirements

• Ollama installed and running
• at least one model pulled
• screenpipe running
• sufficient RAM (8 GB minimum, 16 GB+ recommended)

privacy & security

• 100% local - models run on your machine
• no telemetry - Ollama doesn’t send data anywhere
• no accounts - no sign-up required
• offline - works without internet after downloading models
• open source - Ollama and models are open source

your screen data and AI processing never leave your computer.

resources

• Ollama official site
• Ollama model library
• Ollama GitHub
• screenpipe pipes documentation