Open Source

Works in progress I’m building in the open — a local-first vault, plus agentic tooling for project knowledge, software factories, and MCP servers.

Public Alphasuchconfig/suchconfig-app

SuchConfig

Your local-first source of truth for configs, secrets, and teams.

SuchConfig is built for developers who need one trusted, local-first place for sensitive project work — passwords, configs, secrets, and the tools that support secure day-to-day engineering. It is the vault-and-collaboration product under the Such umbrella: project-scoped knowledge, secrets with placeholders, and credentials — offline by default, merge-friendly over time with CRDTs, with no cloud vault or sync of your note contents.

SuchConfig is an open-core CE built for that: Project Vault + Secrets Vault with encrypted archives, Trusted Folder backup to a folder you choose, and opt-in LAN Wi‑Fi sync as that layer matures. Public Alpha is under heavy active development — APIs, UX, and sync behavior will change.

Local trust — configs and credentials stay encrypted on your machine; no SuchConfig-hosted vault database.
Trusted Folder — user-owned encrypted backup / merge to iCloud, Dropbox, NAS, or USB.
Opt-in LAN P2P — same-LAN sync without a cloud vault; pairing and Handoff verified, incremental deltas still in progress.
Auditable open core — vault crypto and CRDT merge in openvault_core; Apache-2.0 CE you can review and run free.

In developmentzanuka/nocciolo

Nocciolo

kernel / core — the durable core of project knowledge that agents inherit.

AI agents start every session cold — re-learning architecture decisions, coding standards, domain invariants, and the why behind the work from scattered docs and tribal knowledge. That wastes tokens, produces inconsistent output, and breaks long-running agentic workflows.

I’m building Nocciolo as a company brain configuration layer: it turns durable project knowledge — ADRs, READMEs, decision records — into a properly configured memory bank agents can retain and recall. By seeding banks like Hindsight, agents inherit shared context instead of rediscovering it every session.

Knowledge extraction — scans repos for durable knowledge while ignoring ephemeral noise.
Contextual seeding — configures missions and directives for memory banks, tuned for software projects.
MCP integration — emits Model Context Protocol snippets to wire memory into Cursor, Claude Code, and other agent harnesses.
Local-first governance — keeps sensitive architecture decisions under team control, versioned alongside the code.

In developmentzanuka/strumentario

Strumentario

instrumentarium — the instrument set for agentic MCP servers.

Building a useful MCP server is still too much boilerplate and too little product thinking. Teams copy a stdio example, never ship a remote multi-client surface with evals or UI, and rebuild the same query / mutate / validate / scaffold patterns for every domain.

I’m building Strumentario as a local-first MCP content toolkit: one schema produces a remote Streamable HTTP server, MCP app widgets across major frameworks, typed scaffolds with a repair loop, and an AI SDK product path over the same tools — with Braintrust golden traces and compile checks as first-class product, not an afterthought.

Remote MCP server — Streamable HTTP with tools, resources, and prompts; Inspector-ready.
Surfaces — MCP app widgets for React and other UI frameworks (ChatGPT Apps path) plus optional Vercel AI SDK / Anthropic product UIs.
Schema → scaffold — typed clients, pages, and studio stubs with a repair loop from one declarative core.
Measurable evals — Braintrust golden tool traces and scaffold compile checks so quality stays visible.

In developmentzanuka/maglio

Maglio

forge hammer — a factory of factories for agentic software production.

Coding agents made writing code easier. The harder problem is a system that can context-engineer, plan, execute, test, review, and explain itself — without collapsing the moment you step away from the chat.

I’m building Maglio as a CLI that turns a declarative factory blueprint into the concrete files that agents already understand: skills, slash commands, stage harnesses, circuit breakers, and handoff packages. It sits next to Nocciolo — durable knowledge in, structured overnight production out.

Declarative blueprints — configure stages, models, and autonomy profiles instead of copying someone else’s private factory skill.
Stage harnesses — plan → tasks → tests → handoff → loop → review → explain, with stop conditions and circuit breakers.
Progressive autonomy — daytime light profiles and overnight full loops from the same blueprint.
Comprehension-first — owner’s manuals and evidence packages so every run leaves understanding behind, not just diffs.