Weaver Next

Thinking branches. Chat does not.

Weaver is a local-first semantic canvas where you and your coding agent can branch ideas, switch views, and review every AI change. It runs inside Codex as a native plugin and beside Claude Code as a live browser canvas.

Early development · macOS · MIT · 中文介绍

Talk through the work on the left. See and shape its structure on the right.

Why Weaver

Deep work rarely moves in a straight line. You compare hypotheses, follow evidence, revisit an earlier idea, merge useful branches, and abandon dead ends. A chat transcript hides that structure; a generic whiteboard cannot safely tell an agent what the structure means.

Weaver keeps the structure itself:

• Branch without losing context. Scene-aware nodes and typed relationships preserve how ideas connect.
• Work with the exact selection. Codex or Claude Code receives the active canvas, focused node, selection, pinned context, viewport, and revisions.
• Review every AI edit. Agents submit auditable ChangeSets and semantic layout plans. You preview, apply, reject, or undo them.
• Project one graph into many views. Use Canvas, Tree, Graph, Flow, Timeline, Board, Matrix, or Table without duplicating the underlying content.
• Keep your data local. SQLite-backed project state stays under /.weaver/.

Work Side by Side

Weaver is built for collaboration without tab-hopping or copy-pasting context. The conversation and the canvas stay visible together: use language for intent, and use the canvas for structure, selection, review, and control.

The result returns to the same canvas as a reviewable ChangeSet or layout candidate. Apply it, reject it, or undo it; the coding agent never silently replaces the graph.

Install

Codex

codex plugin marketplace add guangtouwangba/weaver-next
codex plugin add weaver-next@weaver

Open a fresh Codex task in your project and ask:

Open the Weaver space for this project.

Claude Code

Requires macOS and Node.js 24 or newer.

git clone --depth 1 https://github.com/guangtouwangba/weaver-next.git \
  ~/.local/share/weaver-next
~/.local/share/weaver-next/install.sh claude

Restart Claude Code or run /reload-skills, then use /weaver-open. Run /weaver-watch to send instructions from the browser canvas.

See the complete installation, upgrade, uninstall, and troubleshooting guide.

The Core Loop

flowchart LR
  A["Build or select ideas on the canvas"] --> B["Ask Codex or Claude Code"]
  B --> C["Agent reads bounded canvas context"]
  C --> D["Preview ChangeSet or layout candidates"]
  D --> E["Apply, reject, or undo"]
  E --> A

1. Start with a goal, a visual template, or an existing Weaver project.
2. Add documents, local images, public links, semantic nodes, and typed relationships.
3. Select the part that matters and ask the coding agent to develop, challenge, reorganize, or lay it out.
4. Review the proposed content or layout before it changes the authoritative graph.
5. Continue exploring, then crystallize the useful structure into a scene-defined artifact.

One Graph, Many Views

A visual template does not create a disconnected copy. It projects the same typed graph into another working structure, with its own layout, theme, and viewport.

Each View remains independently named, pinned, searchable, restorable, and reversible while the underlying content stays authoritative.

What Ships Today

The latest product boundary is documented in the PRD. Roadmap items are not presented as current functionality.

Built for Trust

The graph is content authority; every View has an independent layout document. Content changes increment graphRevision. Moving, pinning, routing, theming, or changing a projection increments only that View’s layoutRevision.

An agent never writes SQLite directly or invents final coordinates. MCP tools validate the active Project, View, Canvas Session, lease, and revision before accepting a ChangeSet or LayoutPlan. A deterministic engine computes layout candidates. Loopback HTTP remains on 127.0.0.1, requires an owner-secret-derived capability token, and recovers dropped events through sequence and revision checks. Preview metadata is written only after an explicit workspace open and is owner-readable only.

Persistent MCP logs and stderr output are off by default. Diagnostics stay in a bounded, redacted memory buffer and never expose preview capability URLs, raw prompts, absolute paths, or stack traces. Temporary logging is explicit, redacted, and bounded; file logging is also owner-only, rotated, and retention-limited.

Develop from Source

Source setup is for contributors, not required for installation:

git clone https://github.com/guangtouwangba/weaver-next.git
cd weaver-next
npm install
npm run build:plugin
npm test

Useful references:

• Product vision
• Using Weaver
• Architecture and agent boundaries
• Contribution workflow
• Release process
• Security policy

中文介绍

思考会分叉，Chat 不会。

Weaver 是一个本地优先的语义画布，让你和 coding agent 在同一张可见、可分支、可回溯的空间里共同研究、规划和写作。它不是让 AI 直接覆盖整张白板，而是让每次内容与布局修改都可以预览、确认、拒绝和撤销。

它解决的是线性聊天难以承载复杂思考的问题：不同探索路径保留各自上下文，同一份语义图谱可以切换为自由画布、思维导图、关系图、流程图、时间线、看板、矩阵和表格。

Codex 用户可直接安装 GitHub Marketplace 插件：

codex plugin marketplace add guangtouwangba/weaver-next
codex plugin add weaver-next@weaver

Claude Code 用户可安装浏览器画布宿主：

git clone --depth 1 https://github.com/guangtouwangba/weaver-next.git \
  ~/.local/share/weaver-next
~/.local/share/weaver-next/install.sh claude

完整步骤、升级、卸载和故障排查见 安装指南。项目处于早期开发阶段，首个公开版本仅承诺支持 macOS。

License

MIT © Weaver contributors.