04 / Open-source tools and attributed adaptation

CreatorStack Systems

A public lead-research demonstration and an attributed cloud-agent runtime evaluation with private production policies.

  • Python
  • Ollama
  • Typed policy contracts
  • TypeScript
  • Vercel Sandbox
Verified proof point

The local lead agent can run from labeled fixtures, while the cloud system keeps outreach draft-only until explicit approval.

Problem

Why this exists

Small operators lose time across research, CRM, website review, follow-up, and fulfillment tools that do not share a repeatable workflow.

Role

What I owned

Workflow design, private policy boundaries, local-model evidence summaries, upstream runtime evaluation, and safety controls.

Original contribution

  • Built a supplied-fixture lead research CLI with a private scoring-policy interface and optional local-model evidence summaries.
  • Added fixture-backed demonstrations without collecting real prospects.
  • Evaluated Vercel Labs Open Agents while keeping CreatorStack production skills and operating playbooks private.
  • Defined draft-only outreach and human-approval boundaries.

System / Architecture

How it works

  1. 01

    The lead CLI validates supplied public-source fixtures and creates CSV and Markdown review artifacts.

  2. 02

    Cloud Agents uses the Vercel Open Agents runtime and sandbox boundary.

  3. 03

    The public fork documents the upstream/private boundary while preserving attribution.

  4. 04

    Artifacts hand work from research to review, approval, and execution.

Limits

What this does not prove

  • Public signals require human verification.
  • The lead agent does not send messages or access private sources.
  • Cloud Agents is an attributed adaptation, not an independently created runtime.

Inspect the work

See the artifact, not just the claim.