Open Source
Four public repositories on GitHub, from an EVE Online intelligence platform to a political agent-based simulation lab, and a trained decision model on HuggingFace.
- Good for
- Code you can read and run: an EVE Online intelligence platform, a news intelligence platform, an agent-based politics lab, a memory server for agents, and a trained decision model.
- Measured
- Over 3,000 unit tests in eve-copilot alone, across 15 of its 17 microservices.
- Two pieces on HuggingFace since September 2026, both Apache 2.0, both with their weak spots written down.
- Open questions
- Who outside the lab runs any of it, and what breaks for them?
EVE Co-Pilot, the largest piece
- 17
- microservices
- more than 600
- tools under the conversational agent
Some of the lab’s work is public. The largest piece is EVE Co-Pilot, an intelligence, industry and fitting platform for the game EVE Online: 17 microservices, a full Dogma engine, live battle tracking from the zKillboard stream, doctrine detection by clustering, and a conversational agent over more than 600 tools. It is alpha software and says so.
The other public repositories
- News MCP. 34 services, an AI analysis pipeline, entity resolution against Wikidata. Before Retina.
- ABM² Digital Lab. Political opinion dynamics, agent-based, in the browser.
- A memory server. Persistent memory for agents.
Before Retina there was News MCP, a 34-service news intelligence platform with an AI analysis pipeline and entity resolution against Wikidata; Retina is what the lab learned from it. ABM² Digital Lab simulates political opinion dynamics with agent-based models in the browser. A smaller one rounds it out: a persistent memory server for agents.
Two pieces on HuggingFace
- laya-rlcd-training. A training loop for the open Laya decision model, which ships the method but not the loop.
- laya-news-decisions. The model that loop produced for news.
Since September 2026 two more pieces are public on HuggingFace: laya-rlcd-training, a training loop for the open Laya decision model, which ships the method but not the loop, and laya-news-decisions, the model that loop produced for news. Both Apache 2.0, both with their weak spots written down.
The author line is the same in every README: Andreas for architecture, design and direction, Claude Code as implementation partner.
On GitHub
Read 4 November 20265 public repositories under one account. Push dates, stars and licences are read from GitHub at snapshot time; the counts of services and tests come from the READMEs.
| repository | what it is | stack | stars | last push |
|---|---|---|---|---|
| cypher-dj AGPL-3.0 | A Linux DJ studio and DAW that a human and an AI agent play together over MCP. Stem separation, live coding, warp. Public beta. | C++ | 0 | 01.11.2026 |
| eve-copilot Apache-2.0 | EVE Online intelligence platform: 17 microservices, Dogma engine v4, real-time combat tracking, doctrine detection, alliance management, a conversational agent over 600+ tools. 17 services · 111+ migrations · 3,029+ tests · 609 MCP tools | Python, TypeScript | 4 | 05.10.2026 |
| persistent-agent-memory | MCP server for persistent agent memory across sessions: semantic search, typed memories. | Python | 0 | 05.04.2026 |
| news-microservices | News intelligence platform: RSS ingestion, multi-stage AI analysis, entity resolution against Wikidata, knowledge graph. The predecessor of Retina. 34 services · RabbitMQ event bus with 12 exchanges | Python, React | 1 | 05.04.2026 |
| abm2-digital-lab MIT | Political agent-based model simulation platform: opinion formation, socio-ecological couplings, emergent macro patterns from micro interactions, empirically calibrated. web-based, FastAPI + React, MCP integration | TypeScript, Python | 0 | 05.04.2026 |
How it is built
Two authors, every time
The author line is the same across the repositories: Andreas for architecture, design and direction; Claude Code as implementation partner. The READMEs call it what it is: human vision and machine implementation, working as equals.
What the public code shares with the lab
The same habits: microservices behind an event bus, an MCP layer so that an agent can drive the system, tests counted and stated, and a status badge that says alpha when it is alpha.