praisonai ollama - PraisonAI

Run a fully local agent against an Ollama model, with Weak-Model-Proof (WMP) execution added automatically to keep small models reliable.

Quick Start

Simplest local run

Ask a question with the default model (llama3.2:3b) and auto-enabled WMP.

praisonai ollama "Explain vectors"

Pick a model

Point at any model you have pulled locally.

praisonai ollama "Summarise this repo" --ollama-model qwen2.5:7b

Disable WMP

Turn WMP off for benchmarking raw output or when the model is already reliable.

praisonai ollama "Explain vectors" --no-wmp

How It Works

The CLI wraps each request in WMP hooks that validate and retry weak-model output before returning the answer. The handler detects model size from the name and selects a matching WMP profile:

Model name contains	Detected size	Profile
`70b`, `72b`, `65b`	`large`	fewer retries
`13b`, `14b`, `7b`, `8b`	`medium`	balanced
anything else	`small`	more retries

Choosing WMP mode

Match the WMP flag to model size and task complexity.

Command options

Every flag comes from add_ollama_arguments in the SDK.

Flag	Type	Default	Description
`--provider {ollama,litellm,openai}`	`str`	—	LLM provider to use
`--ollama-model`	`str`	`llama3.2:3b`	Ollama model name (e.g. `llama3.2:3b`, `mistral`, `qwen2.5:7b`)
`--ollama-host`	`str`	`$OLLAMA_HOST` else `http://localhost:11434`	Ollama server host
`--weak-model-proof` / `--wmp`	flag	auto for Ollama	Enable Weak-Model-Proof execution
`--no-wmp`	flag	—	Disable Weak-Model-Proof execution
`--wmp-strict`	flag	`False`	Stricter validation, more retries
`--wmp-fast`	flag	`False`	Fewer retries, faster execution
`--step-budget`	`int`	`10`	Maximum steps for WMP execution

Environment variables

Local Ollama needs no API key; only the server URL matters.

Variable	Default	Description
`OLLAMA_HOST`	`http://localhost:11434`	Ollama server URL used when `--ollama-host` is not given

--ollama-host always overrides OLLAMA_HOST:

praisonai ollama "Explain vectors" --ollama-host http://192.168.1.10:11434

Common patterns

Fast local chat on a capable model:

praisonai ollama "Draft a release note for v2.0" \
    --ollama-model qwen2.5:7b \
    --wmp-fast

Give a small model more room to work through a harder task:

praisonai ollama "Plan a 3-step migration" \
    --ollama-model llama3.2:3b \
    --wmp-strict \
    --step-budget 20

Use the same local model inside an agent from Python:

from praisonaiagents import Agent

agent = Agent(
    name="Local Agent",
    instructions="Answer using the local Ollama model",
    llm="ollama/llama3.2"
)

agent.start("Explain vectors")

Best Practices

Pin a model per project

Set --ollama-model explicitly so runs are reproducible instead of relying on the default llama3.2:3b.

Keep the server running

Start ollama serve (or the desktop app) before running the command — the CLI exits with a “Make sure Ollama is running” hint if the server is unreachable.

Prefer --wmp-fast on 7B+ models

Larger models rarely need many retries, so --wmp-fast cuts latency without hurting reliability.

Raise --step-budget only when needed

The default of 10 steps handles most tasks; increase it only for genuinely multi-step work to avoid wasted local compute.

Ollama as a model provider

Run agents against Ollama with llm="ollama/…", env vars, and YAML.

Background CLI

Run long agent jobs in the background.

​Quick Start

​How It Works

​Choosing WMP mode

​Command options

​Environment variables

​Common patterns

​Best Practices

​Related