ProofEditor

2026-05-14 · page · source summary

Proof is a free, no-login collaborative document editor built by Every for the specific use case of agents and humans writing together. Agents suggest edits (comments, inline suggestions), humans review and accept or reject. Every character is provenance-tracked to its author.


Key Quotes

"Fast, free, and no login required."

Commentary: The zero-friction onboarding is the product strategy. Unlike Mist (which targets humans collaborating on Markdown) or Google Docs (which is structurally unable to treat agents as first-class participants), Proof is designed from the ground up for the human-agent dyad.

"Every character tracks who wrote it."

Commentary: Provenance is Proof's killer feature. In a world where agents generate 90% of document content, knowing which human wrote which sentence becomes the scarce signal. This is the document-level equivalent of git blame, and it's structural (built into the CRDT) rather than bolted on.

During setup, Proof asks: when should new docs be opened in Proof? Options: all markdown docs; only collaborative docs like plans/specs; only when explicitly asked.

Commentary: This triage prompt reveals the product's self-awareness. Proof knows it shouldn't own every document — only the ones where agent-human collaboration is the point. This is better segmentation than most agent tools manage.

Key Themes

#tool #collaboration #agent-human-interface #provenance #agentic-coding

Proof occupies a specific intersection: it's for documents that are produced by agents, consumed by humans. That's the reverse of most agent interfaces (where humans type instructions, agents read them). In Proof, agents write the first draft and humans edit.

This inverts the typical workflow. In Agent Coding Workflow, the human writes a spec and the agent implements. In Proof, the agent writes the document and the human reviews. Both patterns share the same architecture: one party creates, the other verifies.

The suggestion-mode UX — agents propose, humans accept/reject/reply — mirrors the approval workflow in weft (agents work, humans approve state mutations) and the progressive trust model in Experience Design for Agents. It's the right default: agents can't unilaterally edit shared documents, only propose changes.

The API is clean and well-documented. The edit/v2 protocol with mutation tokens and idempotency keys suggests the team understands distributed systems — this is a CRDT-backed editor, not a simple text field with an API slapped on.

Critical Analysis

What's smart: Proof understands that the bottleneck in agent-human collaboration isn't model capability — it's the interface. Agents can already write good documents. What's missing is a surface where humans can efficiently review, annotate, and approve agent-generated content. Proof fills that gap with a thoughtful permissions model (agents suggest, don't execute) and structural provenance.

What's underrated: The local macOS bridge (localhost:9847) is a genuinely good idea. It means Proof works as a native app with local file access when you want it, and as a web app when you're sharing with remote collaborators. Most tools pick one mode; Proof does both.

What's missing: No version history, no diff view between agent and human edits, no export to structured spec formats. The document model is flat Markdown — which is simple but limits what agents can express. Compare with The Plan Is the Program, where the plan artifact has rich structure. Proof's documents are prose-first, which is right for memos and briefs but weak for implementation plans that need checkboxes, status tracking, and structured metadata.

The Every connection: Proof is built by Every (every.to), a media company that also builds software. This is unusual — most media companies don't ship developer tools. But Every's writers are power users of AI-assisted writing, so Proof is likely scratching their own itch. The question is whether a media company's internal tool generalizes beyond their use case.

Comparison to poietic: Both care about legibility of human-machine collaboration, but from opposite directions. Poietic is heavy infrastructure (dependency graphs, Rust, MIT license, public-benefit corp). Proof is light interface (share a URL, start editing). Poietic asks "how do we coordinate complex work?" Proof asks "how do we write a document together?" Both are right that the answer involves making contribution visible.

Bottom line: ProofEditor is the first tool I've seen that treats agents as first-class document collaborators rather than bolting agent access onto a human-first editor. The provenance tracking, suggestion-mode permissions, and API-first design are the right primitives. Whether it becomes essential depends on whether agent-generated documents become a distinct category of artifact — if agents are writing your PRDs, strategy docs, and research briefs, you want a specialized review surface for that content. If agents are just another set of fingers on the keyboard, Google Docs is fine.


Sources: Proof — Agent-first collaborative document editor Last updated: 2026-05-14