Jerith AI contributes to the open AI ecosystem — OpenClaw skills and plugins kept compatible with the latest releases, distributed compute and storage layers underneath them, and the Jerith projects that put those pieces to work. The public repositories are only the part you can see; a much longer list runs behind the scenes.
Everything below is published on ClawHub under @paradoxfuzzle. I run these on my own unattended agent stack, so they are maintained rather than dumped: when OpenClaw's plugin and skill surfaces move, the packages get updated to match, versions get bumped, and the compatibility work happens before it breaks someone else's deployment. Each card links to its registry entry; source is on GitHub where the repository is public.
Hot/cold context compression before dispatch. Recent turns stay verbatim; older ones are compressed while anchors — paths, commands, numbers, code — survive intact. Fewer tokens, same facts.
Compacts noisy exec/bash tool output before it re-enters the session: strips ANSI noise, parses command output, keeps what reasoning needs. The opt-in complement to Context Squeeze.
Least-privilege interface for user data, personas, and config snapshots in MySQL — input validation and secret redaction baked in, not bolted on.
JER wallet and chain client for OpenClaw: local encrypted wallet, balances, indexed history, on-chain sends with limits, fee estimates, deposit watching, backup/restore, node-side mining. Never invents a balance or a txid.
Image generation through Cloudflare Workers AI (FLUX on the free tier), with optional prompt enhancement through Ollama or the primary agent. Assets on demand, zero cost inside the daily quota.
Builds playable browser games (2D canvas or 3D Three.js) from a brief, then proves it: a screenshot verification loop with luminance and motion checks before delivery is allowed.
Tracks cost, cycle, renewal date and auto-renew, totals month and year, and flags renewals inside a configurable window. Explicit, tested cycle math — past-due is flagged, never clamped. Local data only.
Photograph a receipt; it stores item, price, and warranty end date, then reminds you before coverage lapses — the boring admin agents are actually good at.
Content persistence against a local kubo node: add, pin, and track a per-file TTL in a ledger; a daily timer unpins and garbage-collects what nothing references.
Canon keeper for world-building and roleplay: indexes a markdown lore vault and answers with sourced excerpts — the actual text, never a guess. Expiry rules retire superseded canon; retcon marks keep deliberate exceptions alive.
Six repositories carry most of the public weight: a proof-of-work chain, the mesh that distributes model inference across machines, the storage and compute fabrics underneath them, the agent harness that drives all of it, and a handheld build that puts the original Build engine on a Miyoo A30. Different languages, one theme — infrastructure that has to survive contact with real hardware.
Synced from the GitHub API when this page loads.
Multi-agent harness for coding, research, and operations: web, desktop, and mobile clients share one runtime, with models, sandboxes, memory, and integrations behind provider-neutral interfaces. Self-hostable, with no hosted vendor required for the core workflows.
Jerith Coin (JER) full node: SHA-256 proof-of-work with retargeting, Ed25519-signed transactions, wallets encrypted at rest with scrypt + TOTP 2FA, gRPC/TLS follower sync with block validation, an exchange RPC surface, and an explorer.
Model-homogeneous P2P compute mesh for MoE inference: nodes running the same model pool RAM, VRAM, and CPU; Ed25519-signed discovery; QUIC expert dispatch with cryptographic commitments and redundant replay checks; metered contribution that settles as JER. Includes a reversible migration for existing OpenClaw stacks.
Distributed compute fabric that pools VRAM, RAM, CPU, and GPU across machines into one layer applications request without knowing the topology. Rust agents, a Go control plane with Raft, gRPC, UCX over RDMA/RoCE or TCP, and client SDKs. Pre-alpha, built in the open.
Unified network storage across Linux machines and SBCs: mergerfs pools each node's drives locally, GlusterFS aggregates those pools into one mount visible everywhere. Ships a menu-driven CLI with health checks, JSON output, Nagios mode, Samba/NFS exports, and a web dashboard.
Native ARMv7 build of Chocolate Duke3D for the Miyoo A30 handheld. The engine's 8-bit frame is expanded through the palette while being rotated straight into a locked RGBA texture, so the game fills the portrait panel instead of sitting letterboxed in a 360x360 window. Built against the device's SDL 2.26.1 libraries in a Debian buster cross-toolchain, with the gamepad mapped to Duke's controls and the whole thing shipped as an installable SD-card build.
Jerith AI contributes to far more projects than the public list shows. Some are internal infrastructure, some are private deployments, some are early enough that publishing them would be noise. The visible repos are the tip.
Runtimes, tool loops, routing, memory, approvals, and the glue that keeps an unattended agent honest — much of it running for months at a time, none of it written for a demo.
Systemd units, tunnels, monitoring, backups, and the unglamorous operational work that turns a stack from "runs on my laptop" into something that stays up.
Indexing, embedding pipelines, ledgers, and the movement of bytes between machines that were never meant to share a filesystem.
Porting, migration scripts, and upstream changes for OpenClaw-based setups — including the parts nobody files an issue about until they break.
Three habits that show up in every project above, public or not.
Counters, replay checks, commitment hashes, screenshot gates, offline chain audits. If a feature cannot be demonstrated, it is not described as working.
Memory, wallets, and compute stay with their owner by default. Moving data across a network is a decision, not an implementation detail.
SQLite, systemd, explicit config, reversible migrations. Excitement belongs in the research, not in the restore path.
I publish as NoodlyPandaStories on ClawHub and ParadoxFuzzle on GitHub — the same person behind Jerith AI. Most of my work sits where an AI agent meets a real system: a shell, a database, a chain, a network of machines that were never designed to be a cluster.
That usually means building the small, unglamorous layer that makes the bigger thing possible — a guard that blocks a payload before the model sees it, a compactor that keeps a context window usable, a mesh that lets two GPUs on different desks serve one model, a chain that pays for the inference it meters.
I care about the ecosystem staying coherent: packages that track the platform, migrations that do not destroy what came before, and documentation that tells the truth about what is live and what is planned.
A clean, fast website without a CMS or a monthly platform bill. I use AI to build and maintain the pages, and you pay one flat price for hosting: $100 a year. The build is free when you sign up for annual hosting.
OpenClaw package maintenance, compatibility fixes, infrastructure work, a project that needs a distributed layer it does not have yet, or an AI-built website for a local business — tell me what you are building and what is in the way.
Messages go to the Jerith inbox and, once SMTP is configured, to email. No newsletter, no third-party tracking, no resale.