This overview writes itself. Every project keeps its roadmap as data in its own repository: the goal, the phases, and what “done” means for the current phase, each item with evidence (a PR, issue or document). An agent factory on my own hardware (Ratatoskr and Habitat) reads that every two hours and puts it here, with the changes in the making right now. An item only counts as done with evidence. Nothing sensitive is on this page: no secrets, no document content, no links into private repositories.
What I'm working on
wordsworth
A sovereign pipeline that turns large volumes of Dutch government documents into a searchable, privacy-safe corpus with layered, audited de-anonymisation.
- Proof of conceptingest, anonymise, search, reversible tokens with grants, the console
- Pilot: Gooise Merenone real Woo publication (415 documents) end to end, for a key user
- Data autonomythe organisation holds the keys: originals encrypted, rights per group, a separate key service, keys exportable (ADR-0008, ADR-0009)
- Productionsurvives a node, NORA-aligned, ready for a first organisation
Done in this phase 3/9
Still to do in this phase
- No untokenised names left (names with particles like ter, van der)
- A redacted PDF for every reversible document in the dossier
- Console for the key user: deleted documents hidden, a dossier home
- Readable topics (named by the language model) and search that shows the matching passage
- Share from the console with layered anonymisation (a grant at a PPL, the redacted and reveal PDF)
- Documents survive a node (storage on two nodes, proven)
In the making:
synthetic-long-corpus proposed,
originals-ingest-switch proposed,
originals-crypto-shred proposed,
console-dossier-home proposed,
reanonymise-particle-name-documents proposed
convergence private
An online soulslike with a real crypto stake and seasons; phase 1 is the playable 2D demo, the economy, the site and the pitch, on devnet only.
- Phase 1: 2D demoGDD with a balanced economy, a playable 2D demo on devnet, website, pitch and deck
- 2D launchthe game in 2D for real players, after the legal questions are signed off
- 3Dthe end goal: the same game in 3D (engine candidate: Godot, decision 0004)
Done in this phase 2/8
Still to do in this phase
- Win condition as a pluggable rule module
- 2D demo playable on devnet with every mechanic (zones, bonfire, death and retrieve, duel, marketplace, wallet)
- Website live in the homelab with the demo and an interest form that collects
- Pitch (docs/pitch.md) plus the deck as an HTML/CSS zip
- risks-legal.md and open-questions.md complete
- Reviewer sign-off on everything; no real SOL or mainnet in the code
In the making:
demo-marketplace proposed,
demo-devnet-wallet proposed,
demo-tiled-maps-client proposed
ratatoskr
The coordination layer of Mark's agent ecosystem: channels where people and agents work together, and the factory that turns a decision in chat into a deployed change.
- Phases 0-3: relay and chat agentsself-hosted relay, notifications, chat agents from one source, chat to habitat
- The factory runs itselfpropose, build, merge and deploy from chat, within budgets, without Mark in the loop
- Ecosystemevery project on the factory, readable through one MCP
Done in this phase 5/7
Still to do in this phase
habitat
A minimal platform on which Claude Code agents work on repositories autonomously, inside explicit, non-negotiable limits: work distribution, execution, isolation, escalation, audit.
- The cageone dispatch, Kubernetes Jobs, git as the hand-over, a guard on every tool
- Role chainsarchitect, builder, reviewer, security with real tests and budgets
- Roles from the handbookevery role and skill from one canonical source, for every project
Done in this phase 2/4
Still to do in this phase
handbook
The hub of the personal repository ecosystem: one map of every repo's docs, the canonical agents and skills, and one MCP to read it all.
- Docs hubevery repo's docs aggregated at build time, under one docs contract
- Agent and skill registryone canonical definition per role and skill, read by every consumer
- Ecosystem MCPone read-only MCP for docs, roles, skills, factory status and search
Done in this phase 3/4
Still to do in this phase
homelab
Infrastructure as code for a three-node Kubernetes homelab, run through GitOps, that the other projects can rely on.
- Cluster and GitOpsthree nodes, ArgoCD, certificates, tunnels
- High availabilitylosing one node loses no data and no service
- Hardenedbackups proven, secrets rotated, monitoring that pages
Done in this phase 2/5
Still to do in this phase
bubbels
One app of play-and-learn games for children of 3 to 10: no ads, no trackers, one €1 purchase that opens every game.
- Bubbels, the first gamethe bubble shooter: 30 levels, pop sounds, offline, playable on the web
- Brand app in the storesone app with a tile per game, Dutch and English, the €1 unlock behind a parental gate, on Google Play and then the App Store
- Learning gamesgames that teach, for children who cannot read yet and for those who can
Done in this phase 6/10
Still to do in this phase
bubbels2 private
Bubbels: a free, ad-free bubble shooter for children as a plain HTML/CSS/JS canvas game, later wrapped for Android
- Phase 1the first working version
Done in this phase 1/1
Still to do in this phase
Everything in this phase is done.
In the making:
grid-core proposed,
pwa-and-smoke proposed,
levels-and-progress proposed,
audio-and-ui proposed