The Companion

Chat on the left, instruments on the right. The Companion is the ·Sum family's signature surface — a slender live window kept beside whichever AI you talk to, bridged to it by MCP — the most broadly useful interface idea I've shipped in years.

It exists to solve a blind spot in the bet the whole family makes. If chat apps with deep access are becoming the primary computing surface, then the conversation is where the work happens — but a conversation is a terrible place to see state. Rosters, indexes, records, what-changed-while-I-was-away: the model can recite them, but recitation scrolls away and goes stale the moment it's printed. The classical fix is to build an app — and now you've built exactly the thing the bet says loses: another surface competing for attention, a second place to live, eventually a second chat.

Two standing surfaces

The Companion is the family's refusal of that fork. The doctrine: a user manages exactly two standing surfaces — their chat and the Companion. Everything deeper is transient, singular, and user-invoked. The Companion itself is deliberately modest: a slender, plain-URL browser window, host-agnostic by design — it sits as happily beside Claude as beside ChatGPT, in any browser, as a popped-out panel or installed as a chrome-less app. Slender is the identity: it's an instrument panel, not a workspace. When you need full-size reading, the escalation opens in one named reader window that every escalation reuses — the third window never multiplies.

The bridge is MCP — and the roles never blur

The division of labor is strict. Your own chatbot holds the MCP connection and does all the writing; the Companion is read-only by construction. It refetches when you focus it and on a slow visible tick, so your agent's writes — and its deletions — simply appear, without ceremony. Reality drives the display. The agent owns the writing; the Companion owns the showing; and the two surfaces never trade places.

The affordance rules are just as strict, and there are only two: pills copy, links navigate. A chip's face is the literal text it copies; underlined text moves you. Copy a chip, paste it into your own chat, finish the sentence in your own words — and your agent performs the act. There is deliberately no send button anywhere in the Companion. We considered one — a quick-send box is the obvious “improvement” — and rejected it, because it would quietly recreate a second chat and route around the agent's judgment. The Companion can show dialogue — in Suminar's seminar view it displays the live record of the conversation itself — but it never joins it. It prepares speech; it never delivers it.

A grammar you can copy

The chips work because the family runs on a three-mark grammar: @ names who, # names where, [[…]] names what. An @handle addresses something that can talk back — a person, or a source agent. A #handle names a container — a sum, a seminar. Double brackets cite a page: you hail an @, you enter a #, you quote a [[…]]. Notation is precision, never a requirement — plain language stays first-class, and “the page about the cats” resolves just fine. But when you want to bind your words to an exact object, the Companion hands you the binding: every chip is a piece of the grammar made copyable. It is deixis for chat — you point at the thing by copying its name.

Why a text grammar at all? Because it is the compromise the present forces. Today's host platforms give an external application exactly one channel into the conversation: text the user pastes. MCP lets the agent reach the product's tools, but nothing lets the product reach the user's composer — there is no way for the Companion to hand your chat a typed reference, no dragging a page into the message box, no host-level binding between a chip and the object it names. Until and unless OpenAI and Anthropic open richer interaction schemes between their platforms and outside applications, the copyable mark is the sturdiest bridge available: notation that survives the clipboard, reads as plain language to the person, and resolves as an exact reference for the agent. It is not ideal, and I would retire it the day the hosts allow something better — a chip you drop into the composer that carries its binding with it. The grammar points at that future; the clipboard is just how it ships today.

The host declined to ventriloquize

Look closely at what Claude says in that second screenshot, because I didn't script it. Asked about a source, the host answers: best to ask it directly rather than me paraphrase secondhand — and then puts the question to the named agent and relays its cited answer. That is the exact behavior Against the Middleman and Who Are You Talking To? argue for — the assistant refusing to absorb another voice into its own — emerging from the product's structure rather than from a rule. When the architecture makes the speaker legible — named agents, real authorship boundaries, verified quotation — even the host prefers attribution over paraphrase. You don't have to defeat the ventriloquist. You have to build the room where ventriloquism stops being the path of least resistance.

The chat performs. The Companion shows. Neither pretends to be the other.

What it demonstrates

The Companion is a small thing — a few hundred pixels of read-only browser window — and that is the point. It shows that chat can stay the one conversation you live in without going blind; that a product can be fully present without becoming a persona or a second place to live; and that the interface stays a contract when each surface's authority is explicit: the agent writes, the panel shows, the person speaks in their own chat and holds the only microphone. Both live products ship it today — Suminar and Mem·Sum are free in beta, Companion included — because beside the chat is exactly where the ·Sum family intends to live (the case study).

Essay — building. The ·Sum family's signature surface, shown live. ← All essays