Choosing an Ollama model

Ollama's catalog and model capabilities change independently of Peekaboo. Treat model names, parameter counts, and memory estimates as discovery hints—not proof that a model can drive the agent. Verify the exact tag you install.

#Match the capability to the command

Peekaboo use	Required Ollama capability	Notes
`peekaboo agent`	Tools/function calling	Required even when the task sounds text-only, because the agent acts through tools.
Image analysis	Vision	A vision-only model can analyze an image but cannot necessarily run the agent.
Visual agent task	Tools; vision if the selected workflow sends images to the model	Tool support and vision support are independent.

Use Ollama's current tools filter and vision filter, then inspect the exact local tag:

ollama list
ollama show llama3.1:8b

The model page is the capability authority. A model that merely produces JSON is not equivalent to one implementing Ollama's tool-calling protocol.

#Smoke-test the selected tag

MODEL=llama3.1:8b
ollama pull "$MODEL"

# A small budget proves the native route and one tool round trip.
peekaboo agent --model "ollama/$MODEL" --max-steps 2 \
  "Read the title of the frontmost window and report it"

The model must return a structured tool call, accept the named tool result on the next turn, and then produce a final answer. If it emits prose that only describes a tool call, rejects the schema, or repeatedly calls the same tool, choose a stronger tool-capable tag.

#Resource considerations

Pick a quantization and parameter size that fits available unified memory or VRAM; otherwise Ollama may evict the

model or respond too slowly for an interactive automation loop.

Longer agent sessions replay conversation and tool history. Prefer a context window large enough for the task and

start a fresh session when old context is no longer useful.

Small models can be fast but less reliable with nested schemas and multi-step planning. Test the actual Peekaboo

tool set rather than inferring reliability from a generic benchmark.

Pulling a model requires network access and disk space even when subsequent inference runs locally.

See Ollama for endpoint precedence, the native /api/chat route, step-budget behavior, and the qualified privacy boundary.