Jonathon's AI Wiki

AI Signal Engine — a Hermes-Driven Agentic News Map

3 min read

Source: raw/Fable_5_Build_-_Agent_Controlled_Signal_Map_Open_Source.md (MattVidPro AI, YouTube build walkthrough — transcript)

An open-source “AI signal engine”: a persistent Hermes agent that scans AI news, reasons about what matters, and POSTs structured JSON signal-events into a separately-hosted dashboard it does not otherwise control. The agent-feeds-dashboard split is the interesting part — a clean, reusable safety boundary for any always-on research agent. Single-creator build (medium confidence); the durable value is the architecture/security pattern, not the specific tool.

Architecture

  • Agent: a Nous Hermes agent on a Hostinger VPS (via Docker Manager) runs persistently, researches AI news, and emits JSON events — it does not control the dashboard, only feeds it.
  • Dashboard (separate host): stores and visualizes nodes + connections. Each day is a new map; nodes carry novelty / interest / risk scores, verification flags, and entity offshoots. Clicking a connection shows an LLM-written “connection insight” between two items. Can intake RSS, export to Markdown, and generate-brief produces an executive summary + creator angle + title ideas + story bundles.
  • Cadence: ~15–30 min per pass, ~20+ signals/run; suggested ~3×/day on a loop. The board gets laggy once heavily populated.
  • Model-agnostic: demoed with GPT-5.5 via OpenRouter through Hermes (hermes config set openrouter API key) — no paid Nous plan required. Docker/VPS is only needed for the always-on deployment; local Mac/Windows works for trying it (start.bat / launch-browser.bat, or the Mac/Linux terminal command).

The security boundary (the reusable bit)

  • Agent on a separate VPS = a cleaner trust boundary than running locally.
  • Two distinct secrets: a user-set app password gates the dashboard UI; a separate ingest token gates the agent’s POST endpoint.
  • Capabilities fenced in the prompt: no destructive actions, no publishing, a visible/toggleable log, and per-claim verification flags. The agent can only emit JSON — it cannot mutate or publish.

This “agent only emits structured data; a separate service owns storage/visualization/actions” split is the transferable pattern: it bounds blast radius for any always-on agent regardless of the specific dashboard.

Key Takeaways

  • A self-hosted agentic news radar: persistent Hermes agent POSTs scored JSON signal-events into a dashboard it can’t otherwise control.
  • Security by separation — agent on its own VPS, dashboard behind a user-set password, agent gated by a separate ingest token, no-destructive-actions / no-publishing + per-claim verification baked into the prompt.
  • Model-agnostic and cheap — demoed with GPT-5.5 via OpenRouter; Docker/VPS only for the always-on version.
  • Output is creator-oriented: auto-brief with executive summary, story bundles, title ideas, and a verification checklist.
  • Durable lesson is the agent-emits-JSON / service-owns-actions boundary, not the specific tool.

Try It

  1. Copy the boundary, not the tool: for any always-on agent, have it emit structured data to a separate service rather than letting it take actions directly; gate the emit endpoint with its own token.
  2. Cheap trial: point Hermes at GPT-5.5 via OpenRouter and run a single pass locally before committing to a VPS + Docker deployment.
  3. Throttle cadence: ~3×/day is enough for a daily map; expect UI lag as the node count grows.

Open Questions

  • Repo URL / license were not captured from the transcript — verify before citing as installable.
  • How are duplicate/near-duplicate signal nodes deduped across daily maps?

Graph View

Backlinks