Blueprints & Marketplace

Looking for the fastest way to ship to Loonie? Publish an advisor pack. Use blueprints when you need custom code.

A blueprint is a versioned, distributable package of agent logic — scenarios, system prompts and orchestration — plus a manifest that says who published it, whether it’s free or paid, and exactly which data it needs. Blueprints always run on the user’s own device, against data the user has explicitly granted.

Tiers

Tier Who publishes Example
Free, in-house The app vendor loonie.core_stress_test — rate shock, job loss, market correction
Free, third-party A bank or advisor, as a brand or lead-gen play A credit union’s first-home readiness check
Paid, third-party A bank or advisor monetizing premium content maple_trust.advisory_stress_test (fictional demo) — five scenarios

Create a blueprint in 60 seconds

pip install -e ".[ollama]"
localagents blueprint init my_bank_stress_test --id my_bank.stress_test --publisher "My Bank"
localagents blueprint validate my_bank_stress_test
localagents blueprint run my_bank_stress_test --config config.yaml

init scaffolds two files:

# blueprint.yaml
id: my_bank.stress_test
name: My Bank Stress Test
version: 0.1.0
publisher: My Bank
tier: free            # or "paid" with a price_usd
category: stress_test
summary: Stress-tests a household against a rate shock and a job loss.
entry: agent:run      # module:function the loader calls
inputs:               # data categories requested — the user must grant each one
  - net_worth
  - liquidity
  - cashflow

# agent.py
from localagents.agents import register

@register("my_bank.stress_test")
def run(harness, snapshot=None, **kwargs):
    agent = harness.build_agent(system_prompt="Only reason from the figures you are given.")
    return str(agent(f"Review this data and summarize risks and resilience:\n{snapshot}"))

A manifest’s inputs are a request, never a grant. Apps built on this pattern (like Loonie) show a consent screen listing each requested category in plain language, and the engine refuses to run — before gathering any data — unless every requested category has been granted. Anything the user didn’t check is never sent. Every run records exactly which categories it used.

Scope What it contains
net_worth Total net worth by liquid, investment and personal-asset buckets
liquidity Liquid balance, essential monthly spend, months of runway
cashflow Monthly income, expenses and net for recent months
diversification Allocation across account and asset types
income_forecast Historical income trend and projection
goal_feasibility Goals, targets, dates and on-track status

Licensing paid blueprints

Paid blueprints unlock with an HMAC-signed license token issued at purchase and verified offline — no phone-home at run time. It proves the marketplace issued the token; it is deliberately not DRM, in keeping with the local-first, no-lock-in philosophy.

Feedback without surveillance

Users keep every run locally, with the blueprint version, so they can compare advice across versions. Publishers get zero automatic visibility. A user may choose to export a single run as a precision-reduced, de-identified bundle (figures rounded, no identity or account IDs) and send it to the publisher themselves — the foundation for opt-in, aggregated publisher analytics.

Security model

Read the full design in the architecture doc and the publishing guide.