0.13.0

New

• AI gate. A new flow-control node that reads the latest reports of the other nodes in its flow, answers your question with the instance’s AI model and decides whether the flow continues. Its report explains the answer with charts, and missing data counts as a No. See AI gate.
• Condition with Yes and No branches. The Condition node is in the node picker and works as if/else: its edges are marked Yes or No, and the No edges run when the check fails.
• Get next batch. A new node that sends a dataset on in batches and never sends the same image twice. See Get next batch.
• Trash. Deleted flows and their data, and data nothing has used for a while, go to the Trash. Restore brings them back, Empty trash frees the space, and the Trash empties itself after 30 days. Deleting a flow lists the data that goes with it and its size, and flow cards show their size. See Flows.
• Buy credits. On OvalBee Cloud, Pro workspaces can buy credits on the Billing page. Purchased credits never expire and are spent only after the month’s plan credits run out. Billing also has plan cards and a Credit activity list with the month’s charges line by line. See Compute.
• Audit log for Cloud workspaces. The owner of a Cloud workspace finds Audit log in the main menu, with that workspace’s events. See Audit log.
• Machine in the Run button. The Run button names the machine and its hourly price, its menu lists the machines the node can run on, and Run only this node is a one-off choice in the same menu. See Runs.
• Prelabel quality for live training. Serve live training with Grounding DINO has a Prelabel quality output: how much of what labelers draw the session already finds. A Condition can use it to decide whether a batch is worth pre-labeling.
• More storage options. Self-hosted instances can store data in Azure Blob Storage and Google Cloud Storage. Bucket routes map a fixed name to a bucket you can repoint later, and can be limited to folders and made read-only. Extra storage is a plan feature. See Platform configuration.
• Flows in MCP clients. In MCP clients that support MCP Apps, the agent draws the flow in the conversation as the canvas shows it, with run state and Run and Stop. The agent can stop running nodes after asking you. See AI agent.

Improvements

• Live training samples only from its own labeling pool, and Prelabel with live training adds a version to its pre-label dataset on every run, so each batch’s pre-labels stay in history. Copy keeps pre-labels with the images.
• Benchmarks are grouped by dataset family, default to the latest version, show the version each run was measured on and load much faster. Class filters and evaluation previews use the ground-truth ontology and its class colours.
• Dataset, model and ontology pickers have a Latest (vN) choice in place of the @latest badge.
• The Labeling and Review Outcome statuses input is now Decision buttons, and the labeling tool shows the buttons in the order set on the node. Labeling and Review cards show only their own work, and public flows show live counts with people anonymized. See Labeling statistics.
• Home leads with Overview, has a redesigned My labeling tasks section with Labeling and Statistics shortcuts, and shows plan limits on Cloud.
• AI chat and a Help menu moved to the sidebar footer. AI chat keeps the whole conversation, folds tool steps into one row, and offers Claude Sonnet 5.5 and the GPT-6 models.
• Random sample keeps its last-run output on the latest sample, with or without a destination.
• The Dataset node completes as soon as the flow reaches it, and its name field is a Same as the dataset name checkbox.
• GPU nodes run on the GPU of the chosen machine and fail with the reason when they can’t use it, instead of falling back to the CPU. A run whose machine is lost mid-way is retried.
• Failed training runs report the actual error, with a hint to lower the batch size when they run out of memory.
• Export as .png includes the whole graph. Detach from group takes a single node out of its group. Nodes can have a No icon style.
• Plans set how many deployments can run at once.
• Deleted AI chats are purged after 30 days, and run and deployment logs after 90 days.

Fixes

• Dropdown fields save the option you pick instead of clearing it.
• Live training accepts a sample after Review has taken the image, and Prelabel with live training warns instead of failing when the session refuses.
• Training reports list the model’s evaluations again.
• A widget hidden on a node stays hidden in templates made from the flow.
• Opening an older run’s artifacts shows that run’s outputs, not the latest ones.
• Renaming an S3 storage provider no longer loses its secret key.
• Reloading a page while the server restarts no longer signs you out.
• Collaborators no longer see an error after switching workspace.

Removed

• The Live Examples tab is gone from the template library. Public flow links still work.
• The Reload all configs and Reload selected config buttons are gone from the flow editor: nodes always use their installed template’s config. See Nodes.
• Flow Versions are available on Pro and Business plans only.
• The Code Executor node is removed.
• Billing no longer shows concurrent run caps, and the credits meter at the bottom of the sidebar is replaced by a credits ring next to the space selector.