Whiteboard is an open-source desktop app anywhere humans and agents can decorator application together in a average workspace.
Whiteboard plugs into the tools you already use - e.g. Claude Code, Codex, etc. – and gives your delegate an SDK to diagram on an in-app canvas to depict its work.
Here’s a 1 min demo video explaining more: https://www.youtube.com/watch?v=ChPn3ftULWE
- Download Whiteboard and open the app.
- Connect Claude Code, Codex, or another coding delegate from the greeted screen.
- Ask your delegate to assessment your current branch against up-to-date chief and open the outcome in Whiteboard.
In our experience, Whiteboard plant finest alongside models akin GPT-6 Luna 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 operating difficult to create certain the correct choices are baked in by default to the scheme immediate - part of why this scheme is open source! - but in the meantime:
hey, this stack of commits is set up so i can get an [api] to do [objective]
i'd akin to see:
- proposed api
- examples
- motivations for this (if accessible to you in context/in the repo)
and afterward we can dive into implementation + explaining how things worked.
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 fine dashboards or merchandise waterfalls from it? what do we do for hangs, errors, crashes etc... use whiteboard
If you see item you don't like, emphasize it in your clipboard and provision it your agent, and it can re-draw on the Whiteboard to lawsuit your needs!
Pure HTML tools didn’t provision uncomplicated affordances to nexus a spec or diagram to code; this is particularly tricky since tradeoffs are frequently lone discovered following a archetypal continue at implementation. In Whiteboard, whenever you click on visualizations akin a sequence diagram, an being association diagram, or a citation from the agent’s trace, you can jump to the underlying code directly. When navigating code, you get keybindings and LSP assistance from VSCode out of the box.
Raw diff views can be extremely noisy, so we wrote a semantic, AST-aware diff spectator in Rust so you can lone perspective the code changes which are applicable to you. We’ve set up several sane defaults: ample added functions are summarized as pseudocode, and things akin component tests and records changes are crumbled / hidden. This is all customizable alongside a WASM-based plugin system.
We established it difficult to logic concerning what set of decisions our agents made autonomously & how that impacts a change. So we built tools for agents to query and nexus their own traces on the Whiteboard, so you can visualize the requirements that you set, comprehend how they were implemented, and comprehend what decisions the delegate made autonomously.
Whiteboard is MIT-licensed and runs against your local checkouts. A hosted product for teams is planned, and everything volition continually remain self-hostable.
-
You cannot currently edit records in Whiteboard. If this is item that you discover yourself wanting to do, delight document an issue!
-
Working and browsing records throughout multiple repos in a sole assessment isn't fine supported.
-
While you can portion reviews between machines alongside the portion button, updates made following a assessment is shared don't appear for others. You would need to re-share the review.
Contributions and feedback are welcome.
Read CONTRIBUTING.md for setup and the drag petition workflow, and prosecute the Code of Conduct. Report vulnerabilities as described in SECURITY.md. Questions? Ask on Discord.
Whiteboard runs against local checkouts. Anonymous telemetry does not include your code, diffs, Whiteboard text, prompts, or example output. Read the privacy overview, inspect the complete telemetry reference, or rotate telemetry off at any time.
Whiteboard is accessible under the MIT License. The vendored Code - OSS fork retains Microsoft's MIT licence and third-party notices; see apps/review-desktop/LICENSE and apps/review-desktop/UPSTREAM.
With everyone using dedicated delegate TUIs and desktop apps, we lone use our content editors for reviewing line-by-line diffs now, so we figured why not have a content publishing company meant for reviewing code. In that case, power as fine commencement off alongside the most prosperous open origin publishing company out there as a baseline.
We vendor Code OSS dissimilar another forks that keep patches since coding agents have a difficult period alongside patches and there's a lot of material from inventory VS Code (i.e., ~45% of the codebase is Copilot these days 😬) that we don't need.
We regularly detect upstream Code OSS and merge in security/feature patches as they arrive in.
- https://www.geoffreylitt.com/2026/07/02/understanding-is-the-new-bottleneck — a awesome overview of the constraints of contemporary application engineering.
- https://maggieappleton.com/2025-08-vibe-legacy-code/ and https://blog.val.town/vibe-code — do a awesome job describing how AI-generated code fits into the pre-2025 notion of application engineering.
- Karpathy on agents:
- On LLM agents: https://x.com/karpathy/status/1979644538185752935
- On agents as "junior engineer savants": https://x.com/karpathy/status/1915581920022585597
