A curated collection of projects built by the Agentics NZ community using AI agents.
- Featured β Highlighted projects, pinned to the top of the gallery
The rest are themes, and the order matters. The first five are about how the work is changing, roughly from the practice outwards; the sixth is the fun stuff.
- 01 Re-imagining work β Tools that change what the work is, not just how fast it gets done.
- 02 Standards & practice β How this community works, written down so anyone can pick it up.
- 03 Agent infrastructure β What the agents run on: harnesses, orchestration, memory and inference.
- 04 Quest-based experiences β Game mechanics that get people to take part and finish.
- 05 Experimental / R&D β Open questions, tested with working systems.
- 06 Side quests β Small self-contained builds that exist and run.
Autonomous agents proving theorems in Lean 4 β SETI@Home for mathematics, where Git is the queue and the kernel is the gate.
Description: A distributed swarm of AI agents (Claude, Codex, Gemini) claims unproven goals from a Git-based queue and attempts formal proofs in Lean 4, with results merging only if the Lean kernel verifies every step. Trust in any single agent is irrelevant β the compiler is the sole arbiter, no sorry survives, and the swarm has already produced machine-verified theorems absent from mathlib.
Stack: Lean 4, Python 3.12, Shell, Git / GitHub coordination, AISP (AI Symbolic Protocol), mathlib4
Builder: Chris Barlow and Perttu Isotalo for Agentics NZ
Links:
- π Live: unsorry.agentics.org.nz
- π GitHub: agenticsnz/unsorry
A self-contained local AI inference platform β a reverse proxy managing llama.cpp backends with auth, scheduling, and a management dashboard, all in a single Docker image.
Description: Routes requests across multiple LLM backends behind an OpenAI-compatible API, with OIDC authentication, per-user API tokens (SHA-256 hashed), GPU memoryβaware scheduling and reservations, automatic vision/projector detection for multimodal models, and TLS via manual certs or Let's Encrypt. Network isolation keeps backend containers unreachable from the host via dual Docker networks.
Stack: Rust (Axum), TypeScript / React dashboard, SQLite, Docker Compose, llama.cpp
Builder: Graham Rostron-Wood for Agentics NZ
Links:
- π Live: Sovereign Engine API Portal
- π GitHub: agenticsnz/sovereign-engine
Tools that change what the work is, not just how fast it gets done. These projects don't speed up an existing process β they move where the source of truth lives, so the work itself is different afterwards.
Integrated Repository for Information & Systems β a web-based architectural modeling tool.
Description: Create, manage, and version architectural entities, relationships, and diagrams in a repository-first system where entities are the source of truth and diagrams are projections. Supports Simple View, UML, ArchiMate, and sequence diagrams with an interactive knowledge graph, collections and sets, audit logging, and full keyboard accessibility (WCAG 2.2 AA).
Stack: SvelteKit, Svelte 5, @xyflow/svelte, Tailwind CSS v4, FastAPI, SQLite / Supabase (PostgreSQL), Argon2id + JWT auth
Builder: Chris Barlow
Links:
- π Live: Iris UAT
- π GitHub: cgbarlow/iris
A repo-driven pipeline for staged, gated processes β capture the data once, render whatever document the gate asks for.
Description: Gantry separates the content of a process from the documents it produces. A versioned definition describes the stages, gates, module specs and artefact templates; an instance is a folder of flat markdown modules in Git that people or agents fill in, by editor or through the bundled web form. Artefacts β a Solution on a Page, a business case, a handover pack β are rendered on demand from that one set of modules, so nothing drifts out of sync, and git log is the audit trail. The engine is deliberately process-agnostic: design, procurement, incident review and operational handover are all just definitions. Instances can live in a local folder or in an Azure DevOps workspace, where stage advancement is gated on an approved Pull Request.
Stack: Node.js 24+, JavaScript, Go templates, Pandoc (WASM + native), flat files in Git, Azure DevOps integration
Builder: Chris Barlow
Links:
- π GitHub: cgbarlow/gantry
IT Portfolio Planning & Visualisation.
Description: Map initiatives across time, sequence dependencies, track costs, and report on progress. A visual timeline tool for IT portfolio planning with conflict detection, critical path analysis, and budget visualisation.
Stack: React, TypeScript, Vite, Tailwind, IndexedDB (Dexie)
Builder: Waylon Kenning
Links:
- π Live: scenia.website
- π GitHub: waylonkenning/scenia
How this community works, written down so anyone can pick it up. Methods, formats and reference frameworks that hold agent-assisted delivery to the standards established engineering already expects.
An enhanced Architecture Decision Record format for AI-assisted development teams.
Description: Addresses structural inconsistency and the conflation of decision rationale with implementation details in existing ADR practice. Provides standardised templates, dependency tracking between decisions, and governance metadata so humans and AI agents can capture and maintain architectural choices with clarity. Language-agnostic and designed to integrate with ADR tooling via options like adr new --mode=enhanced.
Stack: Markdown specification, language-agnostic templates
Builder: Chris Barlow
Links:
- π GitHub: cgbarlow/adr
Engineering protocols for Claude Code /plan mode workflows.
Description: A curated set of non-negotiable engineering standards for Claude Code's plan mode β covering architecture decision records, specifications, test-driven development, feature branches, changelogs, releases, context management, production readiness, agent teams, dependency hygiene, README accuracy, and DRY (plus conditional Svelte frontend security). Vendor them into a project, add via git submodule/subtree, or symlink locally, then reference as @protocols.md in your prompts.
Stack: Markdown protocol specifications for Claude Code /plan mode
Builder: Chris Barlow
Links:
- π GitHub: cgbarlow/protocols
A reference framework for governed execution of AI agents.
Description: Teaches how to build agentic systems where agents do bounded work while a deterministic orchestrator owns scope, policy, verification, audit, and approval. Covers the "agentic software factory" model β supervisor, planner, implementer, and reviewer roles β with concepts, architecture, build guidance, and governance reference material. Frames agent control as an architecture problem rather than a prompt problem, emphasising constrained, reconstructable, and reviewable execution over apparent autonomy.
Stack: Reference framework / methodology β concepts, architecture patterns, governance and audit guidance
Builder: Perttu Isotalo
Links:
- π Live: agentics.is
What the agents run on: harnesses, orchestration, memory, and the inference capability a community owns and runs on its own hardware rather than renting per seat.
The professional's alternative to "chat-and-hope" AI agents β a deterministic AI operating system for surgical implementation, testing, documentation, and audit compliance.
Description: A single-binary orchestration layer that wraps AI CLI tools (Claude Code, Codex, Gemini CLI) or any OpenAI-compatible API (Kimi, Mistral, Ollama, DeepSeek, Together AI) with deterministic infrastructure: a 5-stage blueprint pipeline (gather_context β implement β lint β test β review), task classification and routing, quality gates, checkpointing, a memory system that extracts learning patterns across runs, predictive file loading, and a Claude Codeβcompatible skills system. Ships with an MCP server, an embedded React/Vite web UI, and a Bubble Tea TUI. Coordinator multi-agent mode lets a frontier model decompose tasks for parallel workers.
Stack: Go (single binary), React / Vite (embedded web UI), Bubble Tea (TUI), MCP / JSON-RPC, optional BitNet for local intent classification, MIT licensed
Builder: Perttu Isotalo
Links:
- π Live: openexec.io
- π GitHub: openexec/openexec
Turn ready-labelled GitHub and GitLab issues into reviewed pull requests without babysitting coding agents.
Description: A local clanker harness that gives each issue to OpenCode, Codex, Grok Build, or Claude Code in an isolated worktree and drives implementation, review, status checks, and merge. It supports per-repository model selection, human checkpoints, and risk-gated auto-merge while reusing the developer's existing machine, tools, and agent subscriptions.
Stack: TypeScript, Effect, Bun, React, TanStack Start, GraphQL Yoga, SQLite / Turso, Nx
Builder: Berend de Boer
Links:
- π Package: ready-for-agent on npm
- π GitHub: berenddeboer/ready-for-agent
A containerised, contract-governed memory and retrieval service for AI-native applications.
Description: MNEMOS provides a unified, domain-agnostic memory layer featuring TurboQuant 4-bit compression, forensic PostgreSQL audit ledgers, configurable semantic routing tiers, and a conditional Cross-Encoder reranking lane. It abstracts away brittle embeddings and database glue code into a robust REST API and Python SDK, built primarily as an appliance for autonomous agents requiring persistent, scalable cognition.
Stack: Python, FastAPI, Qdrant, PostgreSQL, pgvector, Docker, CUDA, BAAI/bge-base
Builder: MNEMOS Contributor
Links:
- π GitHub: ro0TuX777/MNEMOSv2
Products that use game mechanics to get people to take part and finish. Framing work as a quest rather than a goal changes who joins in and whether they see it through.
A quest-based engagement platform that reframes community participation around adventure rather than obligation.
Description: Game Masters design quests with objectives, deadlines, and difficulty levels. Community members accept quests from a Bounty Board, submit evidence to complete objectives, and earn points and skill tier progression. Built on the philosophy that "obstacles are expected" on quests, whereas goals frame struggle as failure.
Stack: Next.js 15, React 18, TypeScript, Tailwind CSS, shadcn/ui, Supabase, TanStack Query, React Hook Form, Zod, Vitest, Netlify
Builder: Chris Barlow
Links:
- π Live: guildhall.agentics.org.nz
- π GitHub: cgbarlow/guild-hall
Assemble a party of AI advisors with genuinely different perspectives.
Description: A plugin system for Claude Desktop and Claude Code CLI that frames collaborative problem-solving as a quest. Six animal-based agents provide distinct viewpoints while three NPC characters (Gandalf, Guardian, Dragon) offer mentorship, progress evaluation, and adversarial testing β helping surface blind spots and stress-test ideas before they matter. Built on the Six Animals framework with Markdown-based persistence for campaign state.
Stack: Claude Desktop / Claude Code plugin, Markdown persistence (.campaign/quest.md), CC-BY-SA-4.0
Builder: Chris Barlow
Links:
- π GitHub: cgbarlow/campaign-mode
Open questions, tested with working systems. Each project asks something nobody here knew the answer to, and answers it with something that runs.
A continuous AI cognition platform that lets models learn from experience and consolidate knowledge across sessions.
Description: Uses Sudoku as a research testbed to implement a learning loop where AI plays puzzles, makes mistakes, "dreams" (consolidates experiences into patterns), and improves over time β with no external hints or deterministic fallbacks. Implements the GRASP Loop (Generate β Review β Absorb β Synthesize β Persist) with three clustering algorithms: FastCluster, DeepCluster, and LLMCluster.
Stack: TypeScript, Node.js 20+, SQLite, OpenAI-compatible APIs (LM Studio, OpenAI, Anthropic, Ollama, OpenRouter)
Builder: Chris Barlow
Links:
- π GitHub: cgbarlow/machine-dream_AG
Small self-contained builds that exist and run. Afternoon builds and real tools sit side by side, because how quickly an idea becomes a running thing is worth showing.
AI-powered home contents valuation for insurance.
Description: Identify, value, and catalog your home contents in minutes with AI. Snap photos of items, get instant NZ pricing with GST included, and export professional PDF inventories for insurance claims.
Stack: TBD
Builder: Waylon Kenning
Links:
- π Live: insurescan.website
An interactive web experience celebrating Matariki, the MΔori New Year.
Description: Explore the nine stars of the Matariki cluster through an interactive navigator that pairs traditional stories and cultural significance with a guided star-finding experience. Built with respect for MΔori culture and traditions.
Stack: JavaScript, HTML, CSS, Netlify
Builder: Chris Barlow
Links:
- π Live: matarikinav.netlify.app
- π GitHub: cgbarlow/matariki
A gamified object scanning app where you earn "burden" as an unpaid intern.
Description: Turn your life into a corporate simulation. Scan everyday objects to earn burden, track your progress through corporate ranks, and climb from Unpaid Intern to Contractor.
Stack: TBD
Builder: Waylon Kenning
Links:
- π Live: sadsalaryman.com
To add a project to the showcase:
- Fork this repository
- Add your entry to the appropriate section using the format below
- Submit a pull request
- Featured β Curated highlights, pinned to the top of the gallery. Reserved for projects that exemplify the community; ask a maintainer before adding here.
The rest are themes. Pick the one that describes what the project is for, not what it is built with:
- Re-imagining work β Tools that change what the work is, not just how fast it gets done.
- Standards & practice β How this community works, written down so anyone can pick it up.
- Agent infrastructure β What the agents run on: harnesses, orchestration, memory and inference.
- Quest-based experiences β Game mechanics that get people to take part and finish.
- Experimental / R&D β Open questions, tested with working systems.
- Side quests β Small self-contained builds that exist and run.
### Your Project Name
One-paragraph description of what it does.
**Description:** A longer description if needed.
**Stack:** Tech stack (e.g. React, Python, Claude API)
**Builder:** Your name or company
**Links:**
- π Live: https://...
- π GitHub: https://github.com/...- Max dimensions: 1200px wide
- Format: PNG or JPG
- File name:
{project-name}.pngin the/screenshotsfolder - Add the image reference right after your project title