mowa

Load your Mowa agents at runtime. Edit the prompt in Mowa, ship it without a deploy.

Install

pip install mowa
pip install "mowa[httpx]"

No required deps. Uses httpx if it's installed, else the standard library.

Use

from openai import OpenAI
from mowa import Mowa

mowa = Mowa()
openai = OpenAI()

agent = mowa.get_agent(
    "support",
    fallback={"system_prompt": "You are a helpful support agent.", "model": "gpt-4o-mini"},
)

reply = openai.chat.completions.create(
    model=agent.model,
    messages=[
        {"role": "system", "content": agent.system_prompt},
        {"role": "user", "content": "Where is my order?"},
    ],
    temperature=agent.settings.get("temperature"),
)

Mowa() reads MOWA_API_KEY and MOWA_BASE_URL from the environment. You can also pass api_key, base_url, cache_ttl (default 60s) and timeout (default 2s).

Versions

mowa.get_agent("support")
mowa.get_agent("support", version="draft")
mowa.get_agent("support", version=3)

Caching and fallback

• First call fetches the agent and keeps it in memory.
• For the next cache_ttl seconds you get the cached copy. No network.
• After that you still get the cached copy right away, and it refreshes in the background.
• If the first fetch fails and you passed a fallback, you get the fallback. It never raises.
• With no fallback, it raises MowaError.

agent.source tells you where it came from: network, cache, stale or fallback.

Warm the cache at startup with mowa.prefetch(["support", "billing"]). Reset it with mowa.clear_cache().

Traces

span.set_attributes(agent.trace_attributes())

Gives {"mowa.agent": "support", "mowa.version": "v3"}, so traces link back to the version that ran.

Raw HTTP

curl https://mowa.dev/api/v1/agents/support/live -H "x-api-key: $MOWA_API_KEY"

Add ?version=draft or ?version=v3. Send If-None-Match with the ETag to get a 304 when nothing changed.