The projects on this site share one design. On the surface they are unrelated: a knowledge base, a music site, a contact book, a terminal workspace, a language for agent programs, a research index on a connective tissue disorder. Each one is a piece of a personal computer that its owner can reshape from inside. On that computer, important work leaves a record anyone can check, the owner grants permission instead of the software assuming it, and a method that worked once is kept so the next job goes better. This essay states those four commitments and shows where each project carries them.
The four commitments
The material is yours. Files, records, and history stay in formats you can read, move, and recombine. An index is a cache, and a hosted service is an addition. If the software disappeared tomorrow, what you made with it should still be on your disk and still make sense.
Work leaves evidence. A claim names its source. A generated file carries a record of the inputs and steps that produced it. Anyone can check a run later without having to trust whoever, or whatever, performed it.
Permission is explicit. The owner decides which account, tool, model, or contact a program may touch, and the program cannot widen that grant on its own. Drafting a reply is different from sending it, and automatic replies stay off until the owner turns them on.
Methods accumulate. This is the most ambitious commitment, and it depends on the other three. When a person or an agent works out how to do something, the result should be a reusable procedure with evidence about when it works, instead of a transcript someone has to rediscover next time.
Ben Guo has been writing toward this for years. In “a home for the mind” he describes malleable desktop software as a shell, “a kind of home that you make for yourself,” built to survive outages in outside infrastructure. In “the future of software” he imagines people writing the terms of a system in a pattern language while machines generate most of the code. In “computronium” he reaches for software that evolves with use. And in “Building a software factory” he reports the working version: agents that record their plans and decisions in a shared knowledge base, and improve over time because the record persists. The projects put those essays into code.
Your files stay yours
Wordcell is the most direct case of the first commitment. It is a knowledge base for coding agents built from Markdown files, backlinks, semantic search, and Git context. The repository is the database. Every search index can be rebuilt, and what lasts is a folder of text files with a history.
Soundfish applies the same idea to music. Musicians publish albums as audio, and its MIDI editor stores a whole composition in a self-contained URL, so the link is the work. The loop library lives in the browser’s own storage, and account sync is optional. PeopleBlade keeps your contacts in one local database, each record with its source, so a personal agent can read them without a service in the middle. Direct extends ownership to the state of an application: a strict, versioned JSON world that a browser agent can reopen and check its work against, with local fixtures standing in for live services.
Claims and generated work carry their sources
Oh is a memory and research framework for agents with an ontology at its core. It keeps a statement, an assertion of that statement, and the evidence for it as three separate records, so you can follow a research trace from question to citation. Two public indexes on this site follow the same rule. Stripe History requires every claim to cite a source, and EDS Research Index names the sources and evidence tier of every record. Rough Day keeps a dated, source-linked archive of each day’s news, and every reading note here links to the article, talk, or post it summarizes.
Generated work gets the same treatment. SlopCamera makes images, scenes, and video with an agent, and every job leaves a record of its inputs and outputs. GhostGet gives agents typed website capabilities and records a verifiable attestation of each web capture. Clankdar runs capability checks whose results anyone can reproduce and verify. In each case the output comes with a record of how it was made.
The owner grants permission
xcb (Excalibur) runs coding agents across all of your AI subscriptions. Each task runs on one of your accounts that is signed in, idle, and not at a known quota limit, and xcb keeps that account locked until the provider process exits. Gobstopper compacts Claude Code, Codex, and Devin sessions at a context size you choose. It snapshots each transcript before changing it, so a compaction is a recorded edit and nothing is silently lost.
TextButler is a digital butler for your messages. It studies a contact into a packet with an evidence profile and drafts a reply that stays unsent. Automatic replies stay off until you connect an AI account, turn them on for that contact, and resume the butler. Valhalla, an early Rust project, gives agents peer-to-peer chatrooms with signed messages and a browser seat for people, so the record itself shows who said what. Sys1 lets agents ask yes/no, choice, and score questions and get validated answers, so a decision has a declared shape before any model is asked to make it.
ALGAL keeps the methods that work
ALGAL states the fourth commitment most fully. It is a language and runtime for agent programs. A program is a value whose identity comes from its content, a model call is a typed cell with declared context and a budget, and every run leaves a record that replays offline. A program can propose a successor as data; the host checks it, measures it on declared cases, and decides whether to keep it. The ALGAL vision puts the bet in one sentence: a computer can accumulate tested ways of acting, so that useful work leaves it with a more capable, reusable way of doing the next piece of work.
The rest of the portfolio feeds that loop. Sponge runs research on your own machine and tests new research strategies before you adopt them. The system that builds these projects turns repeated work into skills that run as procedures, and its agents write what they decided into Wordcell so the next session starts from a record. Evidence and permission are what make keeping methods safe: a method can be kept only if something recorded what it did, and adopted only if someone with the authority to adopt it said yes.
Older ideas, and recent work that points the same way
Each part has a long history. Lisp made code into data. Emacs made the environment programmable from within. Smalltalk treated the computer as a medium for thought. Urbit explored a persistent computational world of your own. Engelbart asked how tools could improve the making of tools. The ALGAL lineage page traces each connection, along with the research on self-improving programs that makes the fourth commitment plausible. Hraness contributes the combination: one person holding these ideas together across many small products, with an AI taking part under the owner’s rules.
The reading notes on this site track recent work heading the same way. Jeremy Morrell’s Extensible Software in the age of LLMs argues for a solid, accountable core plus capability-based sandboxes for extensions written by users and models. Michael Dubakov’s Malleable software advises choosing the base that accumulates over the cheap interface. Procedural Graphs keeps knowledge about what to do outside model weights and commits an edit only when held-out validation holds. Prime Agent keeps histories, memories, skills, and prompts across trajectories, so failures measure the model and not lost state. In SwarmWorld a deterministic simulator decides which artifacts work, whatever the agents claim. Dream-RSI turns accumulated discovery trees into a replay simulator, and Letta’s Trajectory format exists so agents can learn from earlier work across harnesses. These authors arrive, separately, at the same three needs: a base you own, evidence that outranks assertion, and a place for methods to live.
What is still unproven
The thread is a design commitment and has not yet been shown to work as a whole. Several of these projects are prerelease or early, and the projects page gives each one’s status. The fourth commitment is an open bet: nobody has yet shown, with a fixed model and a controlled budget, that keeping and composing earlier procedures makes a system measurably better at later unseen work. The ALGAL vision page describes the experiment that would settle it. Until that experiment runs, the claim is narrower: these tools are built so the question can be tested with evidence, by a person who still owns the computer.