I Built an Autonomous AI Outreach Swarm — Now It Reports to Notion in Real-Time

#notion #mcp #ai #python

My submission for the Notion MCP Challenge

What I Built

NEXUS Ultra is an autonomous AI agent swarm that runs 24/7 on my local machine, hunting for high-intent developer conversations on Reddit, drafting targeted outreach copy, and scoring its own output — all without cloud APIs or human prompting.

Every cycle the swarm:

Scrapes Reddit for threads matching specific pain signals (agent debugging, LLM observability)
Runs 8+ specialized agents (SCOUT → COMMANDER → COPYWRITER → VALIDATOR → REWARD) in sequence on a shared blackboard
Produces scored, paste-ready Reddit replies targeting real conversations
Saves deployable copy to a local JSON queue for human review

The problem: I had no visibility into what the swarm was actually doing overnight. Logs are overwhelming. The JSON file is opaque. I couldn't quickly see which cycles scored highest, which agents were winning, or which outreach was ready to post.

Notion MCP fixed this. Now every completed swarm cycle automatically appears in a Notion database — score, MVP agent, outreach copy, target thread context, and posted status. Notion became the swarm's mission control dashboard.

How Notion MCP Is Used

I built nexus_notion_reporter.py — a side-car bridge script that:

Reads nexus_deployable_copy.json (the swarm's output file) every 60 seconds
Compares against a local state file to find new cycles not yet logged
Creates a Notion database row for each new cycle via the Notion API with:
- Cycle ID — unique swarm cycle identifier
- Score — REWARD agent's 0–1 quality score
- MVP Agent — which agent produced the best output
- Type — outreach type (REDDIT_REPLY, DM, etc.)
- Posted — checkbox, updated when human posts the copy
- Timestamp — exact cycle completion time
- Scout Context — what thread/signal was targeted
Embeds the full outreach copy as a block inside each Notion page

The Notion database schema is auto-created on first run — no manual setup needed.

Architecture

[nexus_swarm_loop.py]  ←— 8 AI agents running continuously
        ↓
[nexus_deployable_copy.json]  ←— scored outreach output queue
        ↓
[nexus_notion_reporter.py]  ←— side-car bridge (60s poll)
        ↓
[Notion: Swarm Cycle Log DB]  ←— mission control dashboard

The bridge is intentionally decoupled — it reads the swarm's output file only, never touching the running swarm process. Zero risk of disrupting the autonomous loop.

Why This Matters

Before Notion, reviewing swarm output meant: tailing multi-thousand-line log files, manually parsing JSON, no way to track which copy was deployed.

After Notion:

Every cycle is a database row — filterable, sortable, searchable
Sort by score to instantly find the best copy to post
Filter Posted = false → immediate deploy queue
Real-time: new cycle completes → Notion row appears within 60 seconds

Results
8 historical swarm cycles backfilled to Notion on first run
New cycles auto-log within 60 seconds of completion
Zero impact on running swarm (pure side-car architecture)
DB schema auto-provisioned — one command to start

python nexus_notion_reporter.py        # continuous mode
python nexus_notion_reporter.py --once # sync once and exit

Code

Code: nexus_notion_reporter.py (Gist)

Key file: nexus_notion_reporter.py

Built with: Python, Notion API (v2022-06-28), Ollama (local LLMs), Redis, SQLite
Swarm product: VeilPiercer — AI agent monitoring for local LLM developers

DEV Community