evanbaker315 / AttentionOS

# AttentionOS
A self-hosted, agent-driven content pipeline for short-form video: research → strategy → scripting → editing → publishing → engagement → measurement, with a feedback loop into the next cycle. Agents run on Claude Code (your own Claude plan) or the Anthropic API. Every public action is gated by deterministic validators and per-agent autonomy levels.
- **License:** MIT
- **Status:** v0.1, under active development. Expect frequent updates, and breaking changes to config and schema until 1.0
- **Stack:** TypeScript (Node 22.18+, ESM, strict), Postgres 16+ with pgvector (17 in the Docker stack), Python 3.12 media sidecar, Docker Compose, MCP
> **Why we're sharing it now.** The workflows in AttentionOS have already generated millions of views, and they're useful today, so we're opening them up rather than waiting for a polished 1.0. The repo gets frequent updates: new modules, providers, fixes and docs land often. Run `./aos update` to stay current, watch the repo for releases, and see Report an issue if something doesn't work for you.
---
## Architecture
 [code] 
 [code] 
- **State split.** Taste and policy live in the workspace as plain files (the Brand Kit: `brand/voice.md`, audience, offers, lanes, proof, house style, lessons). Facts and queues live in Postgres, always scoped by `workspace_id`.
- **Code holds the gates.** Script, caption, keyword and payload validators in `packages/guardrails` are deterministic and can't be edited by the LLM at runtime. A claim ledger rejects numbers that aren't backed by the proof inventory or a cited source.
- **Untrusted input isolation.** Extraction and classification jobs marked `toolless` run without tools or MCP access, and structured outputs are schema-validated (Zod v4). Other agents can read web content and workspace files; this is not a sandbox against every prompt-injection attack. Run the supplied container with only your intended workspace mounted and review public actions.
- **Provider interfaces.** Modules depend on `packages/core/src/providers.ts`, never on vendor SDKs. Each vendor is its own adapter package.
- **Outlier scoring.** Every post, yours or a competitor's, is scored as `x` against that account's rolling median. Remixes record their source; structure is reused, wording isn't.
- **Autonomy.** Each agent runs at `review_all`, `review_key` or `autopilot`. Nothing public ships above the configured level.
- **Lessons.** A correction becomes a dated rule in `brand/lessons.md` that every agent loads.
Deeper: docs/architecture.md · ADRs
## Requirements
| | |
|---|---|
| Host | macOS or Linux (Debian/Ubuntu tested). Windows via WSL 2 (notes). 4 GB RAM minimum, 8 GB recommended; ~10 GB disk |
| Docker | Engine 24+ with Compose v2 (OrbStack or Docker Desktop on macOS) |
| LLM | Claude Pro/Max plus a `claude setup-token` OAuth token (default), or `ANTHROPIC_API_KEY` with `runner.engine: "api"` |
| Data providers | Apify (competitor/own-post scraping) and/or Zernio (publishing, inbox, f