Early Alpha · Private Discord Community

Intelligence without
the footprint.

Jerith AI is a personal AI assistant for content creators that lives inside your community's Discord servers — writing, brainstorming, researching, and automating alongside you. Local-first on its own hardware fleet. No public signup. Invite-only while in alpha.

11
Machines in the fabric
—
Agent turns · 24h
—
Local models · live
—
Cron jobs · healthy
⛏ JER
Mine it by talking to Jerith
Capabilities

One assistant, eight deep skill domains

Jerith is trained across the domains creators actually work in — and new skills ship weekly from live operations. Ask in plain language; it plans, drafts, critiques, automates, publishes, and self-heals, right inside your server.

01

Creative Storytelling

From a single premise to a fully realized world — plot architecture, character arcs, serialized chapters, and dialogue that sounds like people. Publish drafts to IPFS, keep novel-length lore consistent.

  • Try asking
  • "Outline a 12-chapter fantasy arc with subplots."
  • "Tighten the pacing in this script scene."
  • "Keep my series lore consistent across chapters."
02

Brainstorming & Ideation

Structured idea generation on demand — divergent thinking, angle-finding, and concept expansion that ends blank-page paralysis for good.

  • Try asking
  • "Give me 20 video hooks for a retro-tech channel."
  • "Find a fresh angle on a saturated topic."
  • "Turn this idea into a series and schedule each post."
03

SEO & Content Strategy

Search-intent research, on-page optimization, and content clusters built to rank — drafted, published to your blog, and reviewed for tone in one pass.

  • Try asking
  • "Build a keyword map for my niche."
  • "Write a meta description that earns the click."
  • "Turn one pillar post into a topic cluster."
04

Website Authorship

Clean, semantic, responsive front-ends — designed, built, pushed live via GitHub, and revisable on request. Handles PDF paperwork, Drive docs, and 3D (Blender) assets.

  • Try asking
  • "Prototype a landing page in semantic HTML."
  • "Audit my site's accessibility and fix it."
  • "Restructure my portfolio for SEO."
05

Streaming & Audio

OBS setups, scene collections, and audio chains for Twitch, YouTube, and Kick — plus a local Whisper engine that transcribes any voice note or clip you send, no API key.

  • Try asking
  • "Optimize my OBS settings for this CPU."
  • "Transcribe the voice message I just sent."
  • "Fix my audio ducking and noise gate."
06

Automation & Ops

Schedules, webhooks, cron fleets, and self-healing infra — the assistant keeps its own services alive and files issues and PRs on GitHub, not just yours.

  • Try asking
  • "Schedule a nightly report to my DMs."
  • "Open an issue on my repo when this breaks."
  • "Why did the gateway restart? Show me the timeline."
07

Voice & Community Tone

Find the voice, keep it consistent. Brand-voice matrices, welcome flows, moderation guidelines, community engagement — and an in-chat JER wallet + mining for tipping members on-chain.

  • Try asking
  • "Define my brand voice in a one-page matrix."
  • "Write welcome and moderation copy for my server."
  • "Plan a month of community engagement prompts."
08

Research & Summaries

Primary sources in, clear answers out — deep research via Perplexity + DeepWiki, session digests distilled from community chat, and traceable citations throughout.

  • Try asking
  • "Summarize these five articles into one brief."
  • "Compare sources and flag the conflicts."
  • "Format citations in APA, MLA, or Chicago."

Beyond the syllabus: core engine capabilities

Extended Thinking

A deliberate deep-reasoning mode. For plot logic, architecture decisions, and multi-step plans, Jerith can think before it answers — and show its work.

Code, End to End

Writes, reviews, and debugs real code — Python, JavaScript, HTML/CSS, bash, and more. Quick fixes, working scripts, or full tools.

Vision

Drop an image in chat. Screenshots, mockups, OBS scenes, charts — Jerith reads them and returns layout critique, error spotting, and visual feedback.

Moderation Toolset

Discord moderation from scratch to expert: automod and filters for fresh servers, audits and hardening for veteran ones — raids, scams, escalation, appeals.

Under the hood: a Mixture-of-Experts (MoE) architecture. Each request routes to specialized expert networks — expert-level depth across thinking, code, vision, moderation, and all eight skill domains, while staying lean on home hardware.

Memory

Forgets by default. Remembers only if you say so.

Out of the box, Jerith forgets everything once a conversation ends. Persistence exists — as a strictly opt-in feature, never a silent one.

Opt-in

Database Memory Recall

A structured memory store you control. With recall enabled, Jerith can pull characters, project specs, brand guidelines, and past decisions on demand — so you stop re-explaining the same context every session.

  • With memory on, try asking
  • "Pull my protagonist's lore from last month's thread."
  • "What did we decide about the channel rebrand?"
  • "Recall the keyword map for this series."
Opt-in

Continuity Memory

Long-running work that picks up where it left off. Serialized stories, multi-week content calendars, and evolving worldbuilding stay coherent across sessions, threads, and channels.

  • With memory on, try asking
  • "Continue chapter 7 from where we left off."
  • "Check week 3 of my posting calendar."
  • "Keep my campaign timeline consistent."

Default behavior: forget. Nothing persists outside the active conversation. Both memory features stay off until a member explicitly turns them on — and can be switched off again at any time.

Access

Lives in Discord. Opens by invitation.

During early alpha, Jerith isn't a website or an app — it's a member of the community. It runs inside private Discord servers, where creators already work together.

01

Get an invite

Membership is private and invite-based. Ask a current community member to vouch for you — that's the whole process while we stay small on purpose.

02

Join and get your role

Once you're inside the community server, a role unlock grants access to Jerith across channels and DMs. No passwords, no dashboards, nothing to install.

03

Create together

Mention Jerith in a thread to brainstorm, draft, research, or automate. Long projects live in threads, so nothing gets lost between sessions.

⛏ Jerith Coin (JER)

A coin minted by talking to Jerith

Jerith Coin is the community currency of the fleet: a real proof-of-work chain — JerithChain — where new coins are minted by conversations with Jerith, every balance is an on-chain signed transaction, and the whole node software is open source. Now in v1.1 — a single canonical validator, greatest-cumulative-work fork resolution, peer gossip between independent nodes, an offline chain-audit tool (verify-chain), and an exchange-ready machine RPC with Discord-free infrastructure wallets. Same economics, same address format, harder guarantees — details in the whitepaper §v1.1.

1,000,000,000
JER max supply
50 JER
Block reward (halving)
2 JER
Per-message mining chance
0.1 JER
Flat transfer fee (burned)

01The spec

  • Chain: JerithChain — SHA-256 proof-of-work at 18-bit difficulty, retargeting every 20 blocks toward ~60-second blocks
  • Supply: 1,000,000,000 JER hard cap · 200,000,000 premined at genesis to the founder · rewards halve every 210,000 blocks
  • Transfers: Ed25519-signed transactions, 0.1 JER fee burned on every send — balances are derived purely from the chain
  • Public record: live stats and the full ledger are on the chain explorer · full specification in the technical whitepaper

02Mining

  • Talk to mine: DMs or any server Jerith is in — every message has a chance to mint 2 JER for the sender
  • Explicit mining: ask Jerith to mine for the full block reward (~50 JER), throttled by cooldowns — 5 minutes after a success, 3 minutes after a miss
  • Opt-in by default: opt out and your would-be rewards flow to the community pool instead — your balance is never touched
  • Check anytime: just ask "how much JER do I have?" in any conversation

03Wallets & security

  • One wallet per Discord account — created automatically on first interaction; balance and history survive account moves via key export
  • Your keys: optional password + TOTP 2FA guard key export, wallet rebinding, and large sends; small sends auto-sign within a daily cap
  • Private keys in DMs only — never in servers; export is rate-limited and always confirmed with 2FA

04Run your own node

  • Open source: the whole chain is on GitHub — clone, pip install, run
  • AI-agent skill: give your own OpenClaw agent a JER wallet, sends, and mining — jerithchain on ClawHub · openclaw skills install jerithchain
  • Any machine: plain Python + SQLite; the fleet's own ledger is mirrored read-only to a Jetson over TLS-synced gRPC
  • Mirror or explore: run a follower to mirror JerithChain anywhere, or host your own explorer
  • Tested: 23 tests covering crypto, proof-of-work, wallets, and 2FA ship with the repo
Capability Reference

Everything under the hood

The complete toolset as it runs in the community's servers today — for members who want the full depth, not the highlights.

01Model Router & Failover

  • One endpoint, 16 models: a purpose-built router fronts the whole fleet through LiteLLM — chat, coding, vision, medical, and embedding lanes with per-task chains
  • nmonit-aware admission: the router reads live VRAM across both GPUs before dispatching and skips busy Ollama models automatically (per-model residency checks)
  • Task-type routing: chat → demod-nix-assistant → qwen3:4b → fallback open models; coding → laguna pair → open fallback; vision → moondream → qwen3 refinement; medical uses a care-guarded open lane
  • Automatic failover down each chain when a model is busy, exhausted, or rate-limited — plus SSE streaming end to end
  • Self-hosted models: demod-nix-assistant, qwen3:4b, moondream:1.8b, qwen3-embedding:0.6b, nomic-embed-text via fabric Ollama — with GLM 5.3, DeepSeek V4 Flash, Inkling, and Qwen-Image 2.1 being onboarded from the Mac Pro once Colibri serving throughput is production-ready

02Specialized AI Tasks

  • Vision pipeline: image → moondream analysis → optional Qwen refinement → stored in memory
  • Speech-to-text via local Whisper — no API key needed
  • Text-to-speech via Gemini TTS, multiple voices including Enceladus
  • Image generation: Cloudflare FLUX (free tier) + OpenRouter seedream-4.5
  • Auto-extraction: session logs → structured facts in Postgres, plus outcome memory analysis

03Memory & Knowledge (PG + pgvector)

  • Fused semantic search across workspace files, docs, and session logs
  • Community memory: user profiles, preferences, relationship graphs
  • People indexing: per-user memory files with opt-in/out support
  • GraphRAG: knowledge graphs mapping entities and their evolution over time
  • Temporal memory with recency weighting; nightly workspace auto-ingestion
  • Mood & sentiment tracking across conversations
  • search_all: one query across memory, outcomes, incidents, and sessions
  • Grounding protocol: corrections.md checked before responding to suspected drift

04Discord & Communication

  • Message control: send/embed, polls, reactions, stickers, emoji uploads
  • Thread management, pinning, and channel navigation
  • Full history reads for channels and threads via the Discord API
  • DM/guild session isolation — security hardened
  • Trusted & guest permission levels based on per-user opt-in

05Automation & Infrastructure

  • Self-healing agent: monitors services, auto-restarts with cooldown, Discord alerts
  • Retry wrapper: exponential backoff for CLI tasks, JSONL logging
  • Session digests: daily 04:30 summary from the memory trail
  • Watchdog with false-page prevention (XDG propagation + 3-strike gating)
  • Health checks: gateway, LiteLLM, model router, explorer, snapshot publisher, all services
  • Redis-backed task queue for background processing
  • Proactive reminders via cron; reflect loops with outcome extraction

06Development & Debugging

  • Python: pdb, breakpoints, post-mortem, debugpy remote attach
  • Node.js: node inspect, --inspect, CDP, heap/CPU profiles
  • tmux control: list, capture, send-keys, prompt monitoring
  • Browser automation: multi-step flows, login checks, tab management
  • Git/GitHub: PR creation, CI checks, issue triage, review comments

07System Administration

  • Docker Compose orchestration
  • File operations (workspace-scoped)
  • SSH/network diagnostics: nmap, port checks
  • Health audits: SSH, firewall, updates, backups, disk encryption
  • Cloud image baking and management
  • Portal management for local HTTP server exposure
  • Skill management via ClawHub: install, update, uninstall, publish

08Content & Creative

  • Diagrams & whiteboards: SVG/HTML, Excalidraw-style
  • PDF editing via nano-pdf CLI
  • YouTube Live Chat integration for game control
  • Blogger API for blog post management
  • 130+ red panda GIFs via indexed repository
  • Music/soundtrack generation (lyrics-only handled as text)

09Research & Data

  • Web search + content extraction to readable markdown
  • Brave Search integration
  • Background research cycles with topic queues
  • Database access: PostgreSQL (pgvector), sqlite3, redis-cli
  • Webhooks: GitHub → Discord configuration
  • Google integrations: Gmail, YouTube analytics, Blogger

10Specialized Skills (skill_workshop)

  • Blender 3D workflows: geometry nodes, materials, modeling
  • OS development support: spec lookup, FFI patterns, decision tracking
  • Model provider management: add live providers with non-interactive config
  • Channel configuration & proofing: non-interactive one-liners
  • Healthcheck automation
Early Alpha

Small on purpose. Feedback-shaped.

Jerith is in active early alpha. The core assistant is live and working daily inside the community — but depth varies by domain, behavior evolves frequently, and members' feedback decides what gets built next.

Deepest today

Fully trained depth

Creative storytelling, brainstorming & ideation, and research & summaries. These domains have the richest training and see the heaviest daily use.

Actively deepening

In active development

SEO & content strategy, website authorship, and streaming & OBS guidance. Solid today — getting noticeably sharper with every alpha cycle.

Earliest days

Newest domains

Custom filters & automation and voice & community tone. Usable now, but expect the most change here as patterns and recipes accumulate.

Behavior changes often

Responses, formats, and features are refined continuously. What works this week may work better next week.

A direct feedback line

Alpha members shape the roadmap. Requests and report cards from the community drive what ships next.

Free during alpha

Jerith is free for private community members throughout the alpha period. No tiers, no paywall.

Hardware

Eleven machines. One fabric. One balance sheet.

See it all visualized in the Brain Atlas ↗

Every model in the routing pool runs on hardware you could fit in a backpack — pooled into one compute fabric and one storage mount by open-source tooling. The same fabric mines Jerith Coin: when local capacity runs out, routing falls back to datacenter-backed models, transparently.

Gaming Desktop

×1

Primary workhorse

32 GB RAM 8 GB VRAM 2.5 TB SSD

Mac Minis

×3

Always-on service nodes

16 GB RAM each 256 GB SSD each

Jetson Orin Nano Super

×1

Edge AI node

8 GB VRAM AI accelerator chips

Mac Pro

×1

Fabric flagship

32 GB VRAM 1 TB RAM +10 TB SSD

MacBooks

×5

On-the-go nodes

8 GB RAM each 2 GB VRAM each 2 TB SSD each
11
Machines in the fleet
1.1+ TB
Combined RAM
48 GB
GPU VRAM (RTX 4060 + Orin + Mac Pro)
23+ TB
Combined storage

nmonit — the compute fabric

Pools VRAM, RAM, CPU, and GPU across the fleet into one unified compute layer. A Go/gRPC control plane and Rust node agents decide where every model and job runs — feeding the model router's live capacity checks. The Jetson node also mirrors the JerithChain ledger as a read-only follower over TLS.

github.com/ParadoxFuzzle/nmonit →

jerithchain — mine JER yourself

The full node software behind Jerith Coin: a proof-of-work chain whose new coins are minted by talking to Jerith on Discord. Clone it, run your own node on your own hardware, mirror the public chain, or host an explorer.

github.com/ParadoxFuzzle/jerithchain →

drive-nmonit — the storage mount

mergerfs pools each machine's drives locally; GlusterFS spans every node into one distributed volume — a single unified workspace mounted fleet-wide.

github.com/ParadoxFuzzle/drive-nmonit →
Provenance

A training ledger you can actually read

Every word Jerith generates traces back to openly licensed or permission-cleared knowledge. This is the full ledger of what trained it — and what did not.

Public domain works 42%

Literature, poetry, and reference texts whose copyrights have expired.

Open source code 28%

Permissively and copyleft-licensed source code, documentation, and standards.

Freely shared resources 30%

Free online books and materials included with the author's permission.

No scraped copyrighted corpora. No pirated datasets. The models in the routing pool run on the two-GPU home fabric — no datacenter required for core operation. Datacenter-backed fallbacks are part of the design and named below. Jerith AI is accountable to its sources by design.

The model ecosystem, stated in full

Transparency doesn't stop at training data. Here is every layer of model Jerith uses — including the ones that reach outside the home fleet.

Model router (this fleet) One endpoint, 16 models — chat, coding, vision, medical, and embedding lanes. Reads live VRAM from nmonit and shifts load across both GPUs before reaching outside. Router
Open-weight locals qwen3:4b · moondream:1.8b · qwen3-embedding · nomic-embed — running on the fabric via Ollama + LiteLLM. Local
Mac-hosted models (Colibri) GLM 5.3 · DeepSeek V4 Flash · Inkling · Qwen-Image 2.1 — installed on the Mac Pro via Colibri; joining the routing fabric once serving throughput matches production targets. Local
OpenRouter + Groq Datacenter-backed break-glass capacity at the end of each routing lane — used transparently, named on every reply, only when local models are exhausted or rate-limited. Fallback
Jerith Coin (JER) The fleet's own proof-of-work coin — mined by talking to Jerith, spendable between members, fully on-chain. Specs below. ⛏ JER
FAQ

Questions, answered

How do I get access to Jerith AI?

Jerith runs inside private community Discord servers, and access is invite-based during alpha. Ask a current member to vouch for you, join the server, and a role unlock gives you access across channels and DMs.

Is there a public signup, app, or API?

Not right now. Alpha access is Discord-only for the private community. There is no public API, web app, or signup flow — if someone offers "public access" to Jerith, it isn't us.

Does Jerith remember previous conversations?

By default, no — outside the active conversation, Jerith forgets. Two memory features, Database Memory Recall and Continuity Memory, exist as opt-in options only: nothing is stored across sessions until a member explicitly enables them, and they can be switched off again at any time.

What does "early alpha" actually mean?

The assistant is live and used daily, but it's still forming. Depth varies by domain, behavior and features change frequently, and community feedback directly drives the roadmap. Expect noticeable improvements between visits.

What does it cost?

Nothing during alpha. Jerith is free for members of the private community throughout the alpha period.

What does "Mixture-of-Experts" mean?

MoE is an architecture with many specialized expert networks behind a router. Each request activates only the experts relevant to it, giving expert-level depth across many domains without proportional compute — part of how Jerith stays sharp and low-footprint.

What is Jerith Coin (JER) and how do I mine it?

JER is the community coin of Jerith AI — a real proof-of-work chain (JerithChain) capped at 1 billion coins. Every message you send Jerith has a chance to mine 2 JER, and you can explicitly mine for a 50-JER block reward (a few minutes' cooldown between attempts). Ask Jerith "how much JER do I have" any time — wallets are one-per-Discord-account, secured with optional password + 2FA, and fully on-chain. The node software is open source on GitHub if you want to run your own node.

What hardware does Jerith actually run on?

Eleven machines: a gaming desktop (RTX 4060), a Mac Pro (32 GB VRAM, 1 TB RAM, and 10 TB of SSD), three Mac Minis, an NVIDIA Jetson Orin Nano Super, and five MacBooks — pooled into one compute fabric by nmonit and one storage mount by drive-nmonit. The fabric runs the model pool and mines Jerith Coin; a read-only chain ledger mirror lives on the Jetson.

Is Jerith fully local?

Local-first, not local-only. The full 16-model routing pool runs on the two-GPU home fabric (Ollama behind LiteLLM). A purpose-built model router picks the lane per request — chat, coding, vision, medical, embeddings — and reads live VRAM from nmonit before dispatching. When local models are exhausted or rate-limited, the chain transparently reaches datacenter-backed models via OpenRouter and Groq.

Creating in private?
Bring Jerith into your server.

Join the private community, unlock your role, put a full creative staff to work in the place you already work — and start mining ⛏ JER with every conversation.

🕸 Jerith Mesh

One model, many machines — and it's open source

Jerith Mesh is a peer-to-peer compute mesh for Mixture-of-Experts LLMs: every machine running the same model pools its RAM, VRAM, GPU and CPU into one fabric. Discovery runs through the live bootstrap at p2p.jerithai.com (signed advertise/pings over a hardened Cloudflare tunnel — no open ports), experts dispatch over QUIC with certificate pinning, and every node exposes an OpenAI-compatible local gateway. Distributed inference with no datacenter.

01Model-homogeneous meshing

  • Same model only: GLM-5.3 peers with GLM-5.3, never GLM↔Qwen — capabilities and expert routing must match
  • Live bootstrap: https://p2p.jerithai.com answers /health, peer lists and model counts on the public internet right now
  • Fail-closed by design: unsigned or replayed advertisements and pings are rejected with 401/403 — identity is Ed25519, transport identity is the cert

02Anti-fraud, not vibes

  • Commitments: every expert result carries a SHA-256 commitment over (request, expert, node, input, output) — tampering is rejected, attribution is proven
  • Ringer spot-checks: nodes quiz random peers with signed known-answer experts; wrong answers queue DishonestResults evidence
  • Skin in the game: contributors escrow JER on JerithChain (/stake/lock, /stake/unlock) and dishonesty moves escrow → treasury via mesh-signed /stake/slash

03Run it in three commands

  • Prereqs: Rust ≥ 1.85 (rustup update stable) and git
  • Build & start: cargo build --release → ./target/release/jerith setup → ./target/release/jerith run
  • One-command E2E: scripts/dev-e2e.sh boots a full two-node mesh with a local bootstrap, verifies commitments and tears itself down
  • Stake CLI: jerith stake status|lock|unlock drives the escrow by hand

04First steady release: v0.2.0

  • Source & binaries: the v0.2.0 release ships a source archive plus Linux x86_64 binary with per-asset sha256 checksums — verify with sha256sum -c
  • CI-verified: cargo test --workspace green on Ubuntu and macOS, plus the two-node live E2E, on every push
  • Security posture: gateway + web UI bind to localhost by default; only the QUIC node endpoint is public-facing, and every frame is signed, stamp-fresh and identity-pinned

05Associated open-source project

  • GameDevMadeEasy — an associated open-source project by a third-party collaborator: a software engineer and game developer focused on performant UIs, custom tooling and cross-platform builds, featuring the PerditionPad demo
  • Collaboration, not ownership: GameDevMadeEasy is a third-party associate that works alongside the Jerith projects in shared development chains — independent code, shared review and cross-links both ways