AI编码
github.com ·
@Hacker News
Show HN:Whiteboard(YC W26)——面向深思熟虑软件设计的开源 IDE
四位作者发布开源桌面应用 Whiteboard,让人类与 Agent 在同一工作区共同设计软件。它接入 Claude Code、Codex 等工具,并给 Agent 提供在应用内画布绘图的 SDK。实现上基于 CodeOSS,点击时序图、ER 图或 Agent trace 片段可跳转到对应代码,并保留 VSCode 快捷键与 LSP;另用 Rust 编写 AST 感知的语义 diff 查看器,默认折叠单测与文档改动,并支持 WASM 插件定制;还提供 Decision Log,让 Agent 查询并关联自身 trace,便于理解其自主决策。作者称 Salesforce、Modal 等公司将其用于架构或规格级变更评审,但未给出可验证数据。
AI推荐理由 Show HN 首发帖,作者自述为开源桌面应用,把编码 Agent(Claude Code、Codex 等)接入应用内画布,并给出基于 CodeOSS、Rust 语义 AST diff、Decision Log 等具体设计;对 Agent 协作与代码评审工具有实际参考价值。局限:正文为团队自述,未提供可验证的评测数据,Salesforce、Modal 使用说法无独立证据,星标/安装量不能证明可靠性。
72
Star 1,602
榜单名次 25
正文
Whiteboard
an open-source canvas for thoughtful software design
Download for macOS ·
Windows ·
Ubuntu ·
Fedora ·
Website ·
Discord
Whiteboard is an open-source desktop app where humans and agents can architect software together in a common workspace.
Whiteboard plugs into the tools you already use - e.g. Claude Code, Codex, etc. – and gives your agent an SDK to draw on an in-app canvas to describe its work.
Here’s a 1 min demo video explaining more: https://www.youtube.com/watch?v=ChPn3ftULWE
## Quickstart
1. [Download Whiteboard](https://install.dev.fast) and open the app.
2. Connect Claude Code, Codex, or another coding agent from the welcome screen.
3. Ask your agent to review your current branch against up-to-date main and
open the result in Whiteboard.
## Guidance
In our experience, Whiteboard works best with models like GPT-6 Sol and Claude Opus 5.5 for their intelligence, cost, and speed tradeoff.
Here are a few example prompts of how to use Whiteboard effectively. We are working hard to make sure the right choices are baked in by default to the system prompt - part of why this system is open source! - but in the meantime:
### For a new API change
> hey, this stack of commits is set up so i can get an [api] to do [objective]
>
> i'd like to see:
>
> - proposed api
> - examples
> - motivations for this (if available to you in context/in the repo)
>
> and then we can dive into implementation + explaining how things worked.
### For a change to add telemetry:
> cna you explain to me the telemetry changes form the newest posthog pr https://github.com/devdotfast/whiteboard/commit/4837e107946e27ebad50c282eb0f2585210d2a35 -- what are we tracking, how can we build good dashboards or product waterfalls from it? what do we do for hangs, errors, crashes etc... use whiteboard
If you see anything you don't like, highlight it in your clipboard and give it your agent, and it can re-draw on the Whiteboard to suit your needs!
## Why does this exist?
### Diagrams that lead to code
Pure HTML tools didn’t provide easy affordances to connect a spec or diagram to code; this is especially tricky since tradeoffs are often only discovered after a first pass at implementation. In Whiteboard, when you click on visualizations like a sequence diagram, an entity relationship diagram, or a quote from the agent’s trace, you can jump to the underlying code directly. When navigating code, you get keybindings and LSP support from VSCode out of the box.
### Semantic diff viewer
Raw diff views can be very noisy, so we wrote a semantic, AST-aware diff viewer in Rust so you can only view the code changes which are relevant to you. We’ve set up some sane defaults: large added functions are summarized as pseudocode, and things like unit tests and documentation changes are collapsed / hidden. This is all customizable with a WASM-based plugin system.
### Decision log
We found it difficult to reason about what set of decisions our agents made autonomously & how that impacts a change. So we built tools for agents to query and link their own traces on the Whiteboard, so you can visualize the requirements that you set, understand how they were implemented, and understand what decisions the agent made autonomously.
## Open source, on your machine
Whiteboard is MIT-licensed and runs against your local checkouts. A hosted
product for teams is planned, and everything will always remain self-hostable.
## Known limitations
- You cannot currently edit files in Whiteboard. If this is something that you find yourself wanting to do, please file an issue!
- Working and browsing files across multiple repos in a single review isn't well supported.
- While you can share reviews between machines with the share button, updates made after a review is shared don't appear for others. You would need to re-share the review.
## Contributing
Contributions and feedback are welcome.
Read [CONTRIBUTING.md](CONTRIBUTING.md) for setup and the pull request workflow,
and follow the [Code of Conduct](CODE_OF_CONDUCT.md). Report vulnerabilities as
described in [SECURITY.md](SECURITY.md). Questions? Ask on
[Discord](https://discord.gg/wYvd2cpMQg).
## Privacy
Whiteboard runs against local checkouts. Anonymous telemetry does not include
your code, diffs, Whiteboard text, prompts, or model output. Read the
[privacy overview](docs/privacy.md), inspect the complete
[telemetry reference](docs/telemetry.md), or turn telemetry off at any time.
## License
Whiteboard is available under the [MIT License](LICENSE). The vendored Code -
OSS fork retains Microsoft's MIT license and third-party notices; see
[`apps/review-desktop/LICENSE`](apps/review-desktop/LICENSE) and
[`apps/review-desktop/UPSTREAM`](apps/review-desktop/UPSTREAM).
## On vendoring Code OSS
With everyone using dedicated agent TUIs and desktop apps, we only use our text editors for reviewing line-by-line diffs now, so we figured why not have a text editor meant for reviewing code. In that case, might as well start off with the most successful open source editor out there as a baseline.
We vendor Code OSS unlike other forks that maintain patches because coding agents have a hard time with patches and there's a lot of stuff from stock VS Code (i.e., ~45% of the codebase is Copilot these days 😬) that we don't need.
We regularly monitor upstream Code OSS and merge in security/feature patches as they come in.
## Influences
-
— a great overview of the constraints of modern software engineering.
- and
— do a great job describing how AI-generated
code fits into the pre-2025 notion of software engineering.
- Karpathy on agents:
- On LLM agents:
- On agents as "junior engineer savants":