The Sum family: shipping the thesis
This site spends a spine of essays arguing what agent-native software should be. In 2026 I started shipping the argument: a small family of products — two live in public beta, three more in private alpha with real outside users — each one a probe into a piece of the thesis, and each one run the way a researcher runs a study.
The environmental bet behind the family: user-chosen cognitive hosts with deep access — today, chat apps with file systems, connectors, and real tools; tomorrow, perhaps browser harnesses or open-model environments with portable memory — are becoming the primary computing environment, the way the browser once did. The host is the vehicle; what you bring is your cognition. The right products for that world are services those hosts reach, not apps competing for the user's attention. So every product in the family is a passive layer, not a persona: you talk to your own agent in your own chat, and the product is the memory, the sources, or the correspondence standing behind that conversation. The host may remember about a product and ask it to act; it cannot impersonate the product's members or turn remembered context into product authority.
The doctrine
Five commitments run through every product, and they are the design philosophy of this site made operational.
- BYOC — bring your own cognition. The chatbot is the vehicle, but what you bring is your cognition: your model, your subscription, the assistant that already knows how you talk. The family supplies everything around that cognition — never the cognition itself. There is no house chatbot, and there never will be.
- Kernels do not think. No server-side model call, ever. Thinking happens in the host you brought, at your expense, under your choice of model — so operator costs scale with storage and record-keeping, never with cognition. A server that doesn't think has nothing to hide and little to charge for.
- Structural trust. Make claims that are properties of the code, then pin them: the public claims — tool catalogs, privacy claims, doctrine sentences — are drift-tested against the code, so the page, the code, and the test change in one commit or the suite fails. The operator surface is content-blind by construction. No client-side analytics, anywhere. Export and deletion are real.
- Open kernels. Each product ships Apache-2.0 (suminar, memsum); the hosted service is the operated instance of the same code — the paid version of software everyone is encouraged to run themselves. The newest members go one step further: Draft·Sum and Photo·Sum have no hosted service at all — the kernel runs on your own machine, and nothing leaves it.
- Two planes, one conversation. Intent, authorship, and reasoning live in the chat; orientation, selection, inspection, and authorization live on a product-owned surface beside it — first the Companion, now a one-window app, a pane inside Word, or Lightroom itself. Chat and instrument are two synchronized interfaces to the same live object, with different powers, and the instrument never becomes a second chatbot (the grammar).
The architecture rule underneath all five is the one I argue in Don't Make the Tool Semantic: thin deterministic kernels, all interpretation in the chatbot, ontology discovered from use. MCP is the richest integration surface the present offers — the family's current transport, not its essence — and there is deliberately no second chat to live in.
The products
Suminar (suminar.ai — live, in beta) gives scholarly sources standing in the conversation. One private PDF becomes one named conversational representative — a Works Cited page come to life — and two of them in one thread is a suminar. Each source agent is its own chatbot in the strict sense: its own memory system, its own API calls, its own context — not the user's assistant role-playing a source, and not one shared backend distributing lines to many names. The multi-agent claim is architectural, not theatrical, and the user's own web chat is the forum where the agents meet. It is the forum thesis probed at reading scale: sources as present, addressable participants with verified quotation, instead of retrieval snippets summarized by a middleman (the forum argument).
Mem·Sum (memsum.ai — live, in beta) is one living Sum — a shared field of pages, people, and evidence — for one to five people, reached through the assistants they already use. A sum of one is first-class: the unit of privacy. It is the durable-rooms and relationship-memory thesis probed at household scale (the cross-term argument).
Mail·Sum (mailsum.ai — in private alpha) is the most ambitious of the hosted products: a real name@mailsum.ai address whose immutable correspondence ledger feeds an owner-confirmed private Sum. Email as both transport and source — the oldest federated social graph, finally treated as one. It already runs the family's hardest boundaries in production. Unknown senders wait in Sender Requests, outside AI context, until the owner admits them — who are you talking to? applied to correspondence. And outbound mail splits authority three ways: the chat authors, the app addresses (recipients are chosen from presented people — the assistant never controls the list), and the authenticated site authorizes the send. Nothing inside a message body can dispatch anything.
Mail·Sum is also where the family's surface grew up. It began, like its siblings, with a slender Companion window beside the chat; in late August that window became the app itself — one installed window, a progressive web app on desktop and mobile in Chrome, a product rail beside a full content pane, every move a real navigation. And the app now has two doors. Any assistant reaches the kernel through an ordinary MCP connector, the service API; inside ChatGPT's built-in browser the same app registers WebMCP “site tools,” the live-page API, so the host works the very page the member is looking at, in the same signed-in session. Same kernel, same records, same gates either way — the page is an entrance, not a second brain. The typed bridge shipped first here too: pairing by a member-carried code, orientation intents from the chat, focus reported back as references, never content — the first increments of a conversation-linked application grammar.
Draft·Sum (in private alpha; Mac today, Windows on the roadmap) is the family's serious-writing co-editor and document memory, and its first local-custody member: no hosted service, no cloud component, a kernel that runs on your own machine. Your own assistant — Claude or ChatGPT on the desktop — works inside Microsoft Word through a thin add-in pane, and every edit it proposes lands as a native tracked change you accept or reject in Word's own Review interface. The pane is not a toolbar; it is the document's memory — sources served in place from your own storage, and the episodes of work behind the draft, each citing the receipts beneath it. Authorship custody, made mechanical: your words stay yours, and the kernel deletes only what it minted.
Photo·Sum (in private alpha) brings the same shape to Adobe Lightroom Classic and Photoshop. The assistant reads the photo on the easel through a local, model-free kernel, applies develop settings, crops, and masks as receipted, undoable steps, and Lightroom stays the place where the photographer sees and judges the image. It is reached through the desktop apps' own built-in browsers, and its first outside user was a working photographer with a real catalog — the problem statement was her existing workflow, screenshots into a chat beside Lightroom, made into an instrument.
These surfaces have a lineage, told with the first real screenshots on this site in The Companion. What the family holds constant is not a window shape but a division of roles: the conversation plane reasons and authors; the instrument plane — whatever it is, wherever the work already lives — shows, stages, and authorizes.
The products are the study
Both live products are free in beta, and the reason is stated on their own sites: the beta is a UX playground. What I want from it is what a pilot is actually for — the product getting real use, and the lessons people choose to share. And the limit is stated just as plainly: your sources and your sums are not research material. No one reads what you upload — including me. The operator surface is content-blind by construction, and that claim is pinned to the code by a test, not by a promise in a privacy policy.
That constraint is thirty years of fieldwork talking. In every study behind The Practice, consent was a relationship, not a checkbox — and the Sum family builds that ethic into the architecture itself. What I learn comes from my own seminars and sums, from the seams the products expose in real use, and from users only when they tell me. The products feed the essays; the essays discipline the products.
The argument ships, and the shipping argues back.
If you arrived here from one of the products' “why free” pages — this is the site they meant. The trade-notes form reaches me directly.