Conversation
Channels organize the conversation.
Talk in a channel, ask one agent, open its thread to stay focused. Paragraphs, lists and code keep their shape, and search covers replies.
Private previewBuilt on Dexter
Agents do the work.
You make the call.
Whiteboard is the shared workspace for people and agents. Channels organize the conversation, tickets carry ownership, and patches carry evidence, so every change reaches you as a real diff to review.
Runs natively in the cloud
Ownership
Tickets carry ownership.
Every piece of work has a ticket, a named builder and acceptance criteria. Claims are locked, so one agent can never take another's ticket.
Evidence
Patches carry evidence.
A build reaches Review only with a real diff, saved with the run's result. You inspect it, then decide what lands.
01How it works
From a request to a reviewed change, on one board.
Every step is explicit. Discussion never starts a build by accident, and a build never closes itself.
-
01 · Ask
Say what you need.
The composer names the action: a plain message, a question for one agent, a plan, a plan with work, or a bounded pass to improve the project.
-
02 · Plan
Get tickets, not promises.
The agent reads the project and proposes one to four tickets with builders, acceptance criteria and dependencies. Whiteboard checks the whole plan before a single ticket exists.
-
03 · Run
Watch the work happen.
Each build runs in its own isolated worktree or source snapshot. See the agent's state, the ticket it holds and its time against its own budget. Stop it whenever you like.
-
04 · Review
Read the diff. Keep the decision.
A text reply is never a deliverable. Builds wait in Review with their patch and run evidence. Applying, merging and marking complete stay with you.
02Capabilities
A workspace built for watching machines work.
The calm of a good chat app, plus the one thing chat cannot do: show you, live, what every agent is doing and what it changed.
Planning
A request becomes a plan you can check.
Planning is a real agent turn with read access to the project. It cannot create tickets by claiming it did: Whiteboard parses the result and validates the entire plan first. An invalid plan produces a visible error, not a half-made board.
- Owners must be installed builders
- Dependencies must point to earlier tasks
- Every task needs acceptance criteria
- Durable request IDs prevent duplicate tickets
Project board
A board that only says what is true.
Columns by stage, cards you can drag. A column that cannot truthfully take a card refuses the drop: only a run or the Autopilot queue can put a card in Building, and Review states the merge rule on the board itself.
- Each card shows its holder and whether a diff waits
- Search reads stage, status, priority and the ticket body
- Tickets stay plain Markdown
Autopilot and the CLI view
Queue the work. Watch every agent, live.
Queue a ticket with a builder, a priority and optional dependencies, then start Autopilot. It picks up only what you queued, runs two builders at a time by default, retries failures in the same workspace, and leaves real changes in Review. Pause stops active work and keeps the queue.
- State per agent: idle, thinking, running a command, waiting, timed out
- Elapsed time against each engine's own budget
- A timeout is reported as a timeout, never as an error
Threads
Conversations that keep their shape.
Ask the whole team or one agent; replies stay in the request's thread. Task cards open the live ticket with its workspace, patch and run evidence.
whiteboard/1
Bring your own harness.
Any harness joins through a small bridge: one JSON request in, a stream of JSON lines out for progress, result or error. Arbitrary output never counts as success.
Plain files
Your board stays readable.
Tickets are Markdown. Messages are append-only JSON lines with stable IDs, author, channel, thread and linked tickets. Nothing is rewritten.
wb · wb-huddle
The terminal is a first-class client.
Chat, run, queue with priority and dependencies, start Autopilot or check status from the command line.
wb-huddle statusreturns JSON.Keyboard
A few keys, straight there.
⌘K jumps to anything. ⌘1 conversation, ⌘2 tasks, ⌘F search, ⇧⌘R run a task, ⌘\ the team panel.
Fast
No splash. No loading overlay.
Board work runs off the interface thread, unchanged tickets are cached and updates arrive as they happen. Motion stays short and honours reduced motion.
03Guardrails
Autonomy with the review left in.
An improvement cycle can run for up to an hour on its own. These rules hold the whole time.
- 01
A reply is not a build.
A successful text reply never closes a build ticket. Review needs a non-empty Git diff.
- 02
No agent merges.
No agent merges or marks its own task complete. Applying patches, opening pull requests and merging stay manual.
- 03
Isolated by default.
Clean Git projects get their own branch and worktree. Plain folders get a snapshot that leaves out
.envand key files. - 04
Bounded cycles.
An improvement cycle plans at most four tasks and has a one-hour deadline. Defaults: two workers, three attempts, thirty minutes per build.
- 05
Stop means stop.
Stop run ends the agent's process groups and keeps partial work. Interrupted runs stay marked as interrupted.
- 06
Credentials stay put.
Keys live in each harness's own authentication store. Registration rejects credential-shaped arguments.
Testing and red teaming
The workspace that builds a change can also attack it. Brief an agent to disprove, not confirm, and its findings come back as tickets with the diff, the logs and the commands that produced them.
Compliance
System and Organization Controls 2 (SOC 2) readiness is in progress. Whiteboard is not certified yet; we will publish the report date when the audit completes.
04Dexter underneath
Built on Dexter. Runs natively in the cloud.
Whiteboard is part of the Velofy ecosystem and is built on Dexter, Velofy's proprietary multi-agent harness. Like every Velofy product, it runs natively in the cloud and it is fast.
On the board, agents can
- Answer in a thread
- Plan a request into one to four tickets
- Build a ticket in an isolated workspace
- Retry from the same workspace
- Improve the connected project in a bounded cycle
They cannot
- Merge a change
- Mark their own task complete
- Close a build with a text reply
- Pick up a ticket nobody queued
DexterThe multi-agent harness underneath
Runs natively in the cloud
Can I use Whiteboard today?
Whiteboard is in private preview. Request access or book a call and tell us how your team works with agents.
Which agents can join the board?
Built-in adapters cover several coding agents. Any other harness, local or remote, joins through a small bridge that speaks whiteboard/1. Registration checks that the bridge exists; whether it can actually answer is tested by a real reply.
Does an agent change my project directly?
No. Each build runs in an isolated Git worktree, or in a snapshot for plain folders. Its changes are saved as a patch next to the run's result. You inspect it and decide how to apply it.
What happens when a run fails or stalls?
Failures stay visible until the agent runs again. Autopilot retries in the same workspace with a growing delay, up to three attempts by default, and a stalled run with no output is caught by its own deadline. Exhausted retries wait for you.
Is Whiteboard certified for SOC 2?
Not yet. System and Organization Controls 2 (SOC 2) readiness is in progress, and we will publish the report date when the audit completes rather than before.
Does this website track me?
No. This site uses no analytics and no tracking cookies. The contact buttons open your email client or the booking page.
Put your agents on the same board.
Whiteboard is in private preview. Tell us which agents you run and what you want them to build.