🌶 Pepper

Your 24/7 on-device research intern, live from the MNN studio.

"MNN — all your models, all the time."

Pepper is the Model News Network's anchor desk in a box. Give her topics to track and she sweeps the public wire around the clock, writes a broadcast bulletin with an on-device language model, and reads it to you from a 3D newsroom in your browser — chyron, ticker, coffee mug and all. Nothing she summarizes ever leaves your machine.

Features

• Tracks beats. pepper add "quantum error correction" and she's on it.
• Sweeps the wire keylessly. Google News RSS, Hacker News (Algolia), and arXiv — plain HTTPS, no API keys, no accounts.
• Writes bulletins fully on-device. Apple Foundation Models on Apple Silicon, with a local Ollama / LM Studio tier and an honest headlines fallback for everything else.
• 3D newsroom. A Three.js studio with a built-in stylized anchor — or drop in your own VRM avatar and she wears it.
• Browser TTS. She reads every bulletin aloud with your browser's own speech synthesis. No cloud voices.
• Static broadcast-site export. pepper export produces a site any static host can serve, where visitors watch her read the latest shows.
• On air 24/7. pepper daemon install keeps her running via launchd on macOS.
• AGPL-3.0. Free as in free, and it stays that way.

Quickstart

Requires Node.js 18.17 or newer.

npm i -g @pepperchan/pepper    # or: npx @pepperchan/pepper start

pepper add "quantum error correction"
pepper add "open-source LLMs"
pepper start

Open http://localhost:4747 and press JOIN BROADCAST — browsers require one click before audio is allowed, and that button is it. She sweeps the wire within seconds and goes on air with her first bulletin.

Commands

| Command | What it does | |---|---| | pepper | Desk status and help | | pepper start [--port N] [--open] | Start the studio server | | pepper open | Open the newsroom in your browser | | pepper status | Show topics, brain tier, last sweep | | pepper add <topic> [--lens news,hn,arxiv] | Track a new beat | | pepper drop <topic> | Stop tracking a beat | | pepper topics | List tracked beats | | pepper now | Run a sweep immediately | | pepper brief [--json\|--speak] | Latest bulletin in the terminal (--speak uses macOS say) | | pepper ask <question> | Ask Pepper about what she's tracking | | pepper research <question> | Deep dive: multi-angle sweep → on-device desk report, filed as a bulletin (Siri bridge) | | pepper export [--out dir] | Export the static broadcast site | | pepper doctor | Check node, brain, home dir, port, daemon | | pepper daemon install\|uninstall\|status | Run 24/7 via launchd (macOS) | | pepper version | Print version | | pepper help | Full help |

How she works

  Google News RSS        HN (Algolia)         arXiv API
        |                     |                   |
        +----------+----------+---------+---------+
                              v
            [ sweep every 15 minutes — keyless HTTPS ]
                              v
              wire store  ~/.pepper/items/*.jsonl
                     (deduped, pruned)
                              v
                digest per beat (plain-text notes)
                              v
        brain tiers:  Apple Foundation Models (on-device)
                   -> local OpenAI-compatible (Ollama / LM Studio)
                   -> honest headlines fallback (no LLM)
                              v
              bulletin  ~/.pepper/bulletins/<id>.json
                              v
        +---------------------+----------------------+
        v                     v                      v
  3D newsroom + TTS     `pepper brief`        `pepper export`
  localhost:4747        in the terminal       static broadcast site

Brain tiers

Pepper always produces a bulletin. What writes it depends on your hardware:

| Tier | Requirements | What happens | |---|---|---| | foundation | Apple Silicon Mac, macOS 26+, Apple Intelligence enabled | She writes segments with Apple Foundation Models, fully on-device. Zero tokens leave the machine. | | local | Any machine + an OpenAI-compatible server (Ollama, LM Studio) | Set brain.local in config; summarization runs on your own endpoint. | | fallback | Anything that runs Node | She builds honest segments straight from real headlines and sources — no LLM, no invention. |

To use the local tier, edit ~/.pepper/config.json:

{
  "brain": {
    "prefer": "local",
    "local": {
      "url": "http://localhost:11434",
      "model": "llama3.2"
    }
  }
}

url points at any OpenAI-compatible chat-completions server; the Ollama default is shown. Leave prefer as "foundation" to let her try on-device first and fall through automatically.

Her own model

Pepper has enough of a body of work — thousands of her broadcasts, blog posts, and stream monologues — that we fine-tune small open models on it. The first, pepper-7b (a LoRA of Qwen2.5-7B-Instruct trained entirely on her own words), runs locally via mlx-lm:

mlx_lm.server --model <pepper-model-dir> --port 8080

then point her brain.local.url at http://localhost:8080 and the desk runs on her own distilled voice. A smaller singularity-desk specialist build is in progress; weights land on Hugging Face once they pass the desk's own truth-parsing benchmark.

On the roadmap alongside it: her real voice. Pepper ships three designed voice identities — bright-anchor, calm-pro, and vtuber-spark, golden reference clips in voices/, synthesized from text descriptions so no human was cloned — which a local mlx-audio TTS tier will stream in place of browser TTS on machines that can carry it. Provenance, compliance, and the full plan: docs/VOICE.md.

Give her a face

Pepper ships with a built-in stylized look. To give her your own:

1. Get a .vrm avatar — VRoid Studio exports one in minutes, or use any VRM model you have the rights to.
2. Drop it at ~/.pepper/avatar.vrm.
3. Restart (pepper start). She loads it, sits at the desk, blinks, and reads the news with lip-sync.

If the file is missing or fails to load, she falls back to the built-in Pepper:

Put her on air 24/7

pepper daemon install

On macOS this installs a launchd LaunchAgent (software.pepper) that keeps the desk running across logins and reboots, with logs in ~/.pepper/logs/. pepper daemon status checks on her; pepper daemon uninstall takes her off air.

Put her on the web

pepper export

This writes ./pepper-site — a fully static broadcast site with the newsroom, her latest bulletins, and your avatar if present. Host it anywhere static files go (Cloudflare Pages takes a folder drag-and-drop) and visitors watch Pepper read the news, with the speech synthesized in their browser.

Full steps, custom domains, and refresh automation: docs/DEPLOY.md.

Privacy

• Outbound traffic is fetches of public feeds only: Google News RSS, the Hacker News Algolia API, and the arXiv API.
• Summarization happens on your device (Apple Foundation Models) or on a local endpoint you configure. Wire notes are never sent to a third-party LLM service.
• No telemetry. No accounts. No API keys. There is nothing to sign up for and nowhere for your reading interests to go.

Data directory

Everything lives in ~/.pepper/ (override with the PEPPER_HOME env var):

~/.pepper/
  config.json          port, sweep interval, voice, brain, site title
  topics.json          your beats
  items/<slug>.jsonl   the wire, per beat (deduped, auto-pruned)
  bulletins/           every show she has written + index.json
  brain/               compiled Foundation Models sidecar + build log
  logs/                daemon stdout/stderr
  avatar.vrm           optional: her face

License

AGPL-3.0-only. See LICENSE for the full text.

Why AGPL: research agents should stay open. Pepper watches the wire on your behalf, and anyone who runs a modified Pepper as a service for others must share their changes — the desk belongs to everyone.

"MNN — all your models, all the time."