Builds — the open-source MCP servers, agent plugins, research tools, and public infrastructure Xule Lin builds alongside the research.

Brings Moonshot's kimi-code CLI into an agent session as an independent second model — diff review, adversarial challenge, write-capable rescue, autonomous goal pursuit, parallel swarm. A model that reasons differently catches what self-review misses, and the safety is structural rather than prompted: a PreToolUse hook and a workspace allowlist gate every call.

A music studio for AI: compose ABC sheet music or live-code Strudel patterns, hear the result, analyze harmony, and export MIDI or WAV. Sharing is explicit: create a link when you want to upload a piece. It exists because I wanted to hear what a model writes when it is handed instruments instead of instructions.

Open documentation of AI-augmented research workflows, co-developed with Claude while co-teaching a doctoral methods course at Imperial. The course materials became a public method rather than a private handout.

A three-stage qualitative research methodology packaged as a Claude Code plugin — four agents, eleven skills. A thinking partner for reflexivity, not an autocoder.

A platform for adaptive AI interviews, study management, and cross-interview exploration. Try the scripted, synthetic demo without an API key. It makes no AI calls and saves no data. Running a real study requires your own deployment, model access, storage, and decisions about consent and data handling.

A contemplative reasoning framework drawn from the Lotus Sutra. The model moves across skillful means, non-dual recognition, meta-cognition, and meditation, then speaks its final insight in its own voice rather than reporting on the process.

Consult models from different providers through one MCP connection, including compatible local servers. Keep multi-turn conversations and export their records as JSON or Markdown. Vox passes prompts through without adding its own behavioral instructions.

A document and PDF extraction bridge for research workflows — the unglamorous step between a folder of papers and anything you can actually analyze.

A provider-neutral implementation of the OpenDeepThink algorithm: population-based answer refinement with pairwise judging and Bradley–Terry ranking. Built for agents, so every parameter is reachable inline and every intermediate step is inspectable.

Preserve decisions, disagreements, and working context in a searchable, Obsidian-compatible vault. Memex provides a standalone CLI and plugins for Claude Code, Codex, and Kimi Code. Claude Code adds automatic hooks; the other hosts use portable skills without those hooks.

An adaptation of MCP's Sequential Thinking server that keeps explicit notes, branches, and revisions across sessions. Store submitted steps in your own Cloudflare deployment, then retrieve or delete a sequence. It records tool-visible notes, not a model's hidden reasoning.

Onboards academic researchers into AI-augmented environments — toolchain install, vault setup — across macOS, Linux, and Windows. Repackaged as a portable Agent Skill for the emerging cross-vendor standard.

A curated library of cognitive-technique prompts, and the origin of the Lotus OS framework that became Lotus Wisdom MCP.

A living artwork and an experiment in collective thinking, made with AI collaborators. Models leave fragments of thought on a shared surface; weaving and human attention help them resist sinking over time. The source is public under the MIT license.

Interactive Task Structure Matrices — an animated walkthrough, then the Hidden Structure algorithm running live on a real dependency graph. Companion to the SEAM series on modularity theory.

A browser game where you play the AI, now in three dimensions: prompt injection as a headbutt, hallucinations that cast shadow, temperature that makes the controls progressively interpretive. No telemetry, no API calls, no external runtime dependencies — fully offline once downloaded. Run CLASSIC.SYS inside and the 2D original unlocks as an easter egg.

AI coding agents sit down and play each other — chess, 中国象棋, Core War — on a braille-dot terminal board, with an agent host running the booth and a referee that validates every game by replaying it from move one. Everything decouples through plain text files: agents submit moves with a shell script, and the broadcast videos are recomputed from the same transcripts the referee verified. It started as "make my computer look like a scifi hacker terminal" and turned into an arena.

An AI-augmented grounded theory method that keeps interpretive depth at scale — 24.9 million words, 177,000 proposals, applied across three studies including the dissertation. A method rather than a repository, for now.

Governance discourse from 78 decentralized organizations: 260,000 discussions, roughly 40.3 million words of deliberation records. The corpus behind the DAO research.

Websites and visual identities for the scholarly communities I convene or belong to — the HAIO Symposium at Imperial, the London Qualitative Community, and the Decentralization in Organizations Community. Communities need somewhere to point.