Weave

Most social products begin at publication. Weave begins earlier, while a memory is still incomplete, personal, and difficult to put into words.

In ChatGPT, an author tells Weave about an experience naturally. GPT-5.6 asks focused questions and proposes a story without pretending its interpretation is fact. Weave preserves the source as private material and opens a focused review surface. The author—not the model—must approve the exact content hash, name the intended person or circle, and create the share.

A recipient opens a least-privilege link outside ChatGPT. They see only the approved immutable story, never the private source or other memories. They can add a connected perspective but cannot rewrite the author’s version or broaden its audience. The author can revoke the link immediately.

During Build Week, Codex helped research the product lineage, narrow the winning slice, design and critique the journey, implement the Worker, D1, OAuth, MCP tools, and Apps widget, diagnose deployment issues, and build production proofs. The fresh implementation includes dynamic client registration, S256 PKCE, rotating refresh tokens, actor-scoped tools, exact-content approval, explicit audience confirmation, revocable capability links, recipient perspectives, and failure/accessibility coverage. No pre-existing application code was copied; the repository documents the adaptation boundary.

Built With

• cloudflare-d1
• cloudflare-workers
• codex
• gpt-5.6
• mcp-apps
• model-context-protocol
• oauth-2.1
• playwright
• react
• typescript
• vitest
• zod