AI Agent
Build, change and debug computer vision flows in plain words, in AI Chat or from your own agent over MCP.
Every night, import new images from our S3 bucket, pre-label them with the PCB detector and send uncertain ones to labeling.
5 steps
Done. Every night at 02:00 the flow imports new images, pre-labels them with PCB detector @latest and sends predictions between 0.25 and 0.6 confidence to labeling. I left Labelers empty: choose who labels before the first run.
Add a check after labeling: images with no objects or tiny boxes go back to labelers, the rest go to review.
5 steps
Added a Labeling gate after Labeling. Images with no objects or with a box under 0.1% of the image go back to labelers, the rest move on to Review. Other nodes keep their settings.
Last night's training run failed. Find out why and fix it.
5 steps
Training stopped before the first epoch: the validation dataset PCB val is empty after last week’s cleanup. I added a Train/validation split of v12 and wired both outputs into Train YOLO.
The agent wants to run a node.
Train YOLO and the nodes after it.
What trained last night, and what should I do next?
4 steps
Train YOLO finished at 03:40 and registered PCB detector v8: mAP 0.681, up from 0.643 on v7. The evaluation report says most misses are small capacitors, so label more of them next, starting with the 87 images the report flagged.
Both ways use the same tools and the same rules, so a flow started in one can be finished in the other.
Open the chat on any page or next to the canvas. Its edits land on the flow while you watch, and you can take over by hand at any point.
Connect Claude Code, Codex or another MCP client with a personal token. It works with your permissions and nothing more.
Examples
A few prompts to start from. Copy one into AI Chat or your own agent and change the names to your datasets and models.
Build
“Create a flow that labels Street scenes, sends every labeled image to review and returns rejected ones to labelers.”
Build
“Split PCB components 80/20, train YOLO and EdgeCrafter on it, and compare both detectors on the validation set.”
Build
“Before labeling, remove near-duplicate images and set dark or blurry ones aside in a separate dataset.”
Modify
“Run the import every weekday at 06:00 and add a note on the canvas that explains the schedule.”
Modify
“Replace Train YOLO with Train DEIM, keep the same datasets and tell me which nodes after it need a change.”
Modify
“Lower the confidence range of Filter predictions to 0.2–0.5 and leave everything else as it is.”
Explain
“Explain what this flow does, which datasets it reads and writes, and what runs on a schedule.”
Explain
“Which nodes can find label errors in a detection dataset, and what does each of them need as input?”
Debug
“The last run of Train YOLO failed. Read the logs, tell me why and fix the flow if you can.”
Debug
“Find and fix missing inputs in this flow.”
Debug
“Which of my flows failed in the last week, and what went wrong in each of them?”
Analyze
“Open the latest evaluation report and tell me which classes the detector confuses most often.”
Built for computer vision
It debugs failed runs, understands how computer vision flows fit together, asks before it runs anything and keeps every change reversible.
Ask what went wrong. The agent checks node states, compares with the last good run, reads the error and log lines, and looks at the data behind each input. Then it names the cause and proposes a fix.
2131,128 images, 6 classes
214Converting PCB val
215Found 0 images
216ValueError: empty val set
FixSplit v12 into train and validation
Editing a flow never starts a run. In AI Chat, every run, stop and revert waits for your approval.
The agent wants to run 2 nodes.
Each edit by the agent, a person or a node run is a step in the flow history. Plans with version history also save a version before every agent edit.
Flow history
The agent follows a detailed guide to datasets, models, labeling and review, so the flows it builds are the ones your team would build by hand.
AI AgentMCP
OvalBee runs a remote MCP server. Add it to your client once, and your agent works with your flows through the same tools as AI Chat.
Connect to the OvalBee MCP server.
URL: https://app.ovalbee.com/api/mcp
Transport: Streamable HTTP
Header: Authorization: Bearer YOUR_TOKEN
Then read the OvalBee guide and list my flows.claude mcp add --transport http ovalbee https://app.ovalbee.com/api/mcp \
--header 'Authorization: Bearer YOUR_TOKEN'[mcp_servers.ovalbee]
url = "https://app.ovalbee.com/api/mcp"
headers = { Authorization = "Bearer YOUR_TOKEN" }{
"mcpServers": {
"ovalbee": {
"url": "https://app.ovalbee.com/api/mcp",
"headers": { "Authorization": "Bearer YOUR_TOKEN" }
}
}
}{
"mcpServers": {
"ovalbee": {
"httpUrl": "https://app.ovalbee.com/api/mcp",
"headers": { "Authorization": "Bearer YOUR_TOKEN" }
}
}
}{
"ovalbee": {
"type": "http",
"url": "https://app.ovalbee.com/api/mcp",
"headers": { "Authorization": "Bearer YOUR_TOKEN" }
}
}Get YOUR_TOKEN on Home in OvalBee under Build with your agent. The dialog fills in your address and token.
AI Chat works through the same tools. Every change an agent writes is validated first and saved as a revision you can roll back.
Create a whole flow from one description, change it with patches, check every change before saving it and step back through history.
flows_listflow_readflow_createflow_validateflow_writeflow_historyflow_revertflow_graph
Search the node library by task or data type and read input and output contracts, valid choices and default configs.
nodes_searchnodes_describenodes_choicesnodes_defaultnodes_categories
Find datasets, models, secrets and ontologies by name, and create the missing ones in one batch.
resources_listresources_create
Run or stop nodes, wait for them to finish, and read node states, past runs, errors and log lines.
flow_runflow_stopflow_waitflow_statusflow_taskstask_readtask_logs
Read node reports section by section, with their findings, charts and data files, and the OvalBee guide for agents.
report_readreport_fileovalbee_guide
FAQ
An OpenAI or Anthropic model chosen by the administrator of your OvalBee instance, such as Claude Sonnet or a GPT model. A conversation stays on the model it started with. If you would rather use a model of your own choice, connect your agent over MCP.
On OvalBee Cloud, AI Chat is billed in the same credits as the rest of the platform, per model call. An agent connected over MCP runs on your own subscription or API key, so OvalBee only bills the nodes it runs.
No. In AI Chat, running, stopping and reverting a flow each wait for you to approve them, and editing a flow never starts a run. Over MCP, the agent can run nodes only if your role allows it, and your client asks before each tool call as it does for any other tool.
Every edit is a step in the flow history, so you can ask the agent to revert it or press Ctrl+Z in the editor. On plans with flow version history, OvalBee also saves a version named Before agent edit, which you can restore from Versions yourself.
The agent writes against the revision it read. If a person, another agent or a node run changed the flow in the meantime, its write is refused with the difference, and the agent reads the flow again before it tries once more.
No. It refers to secrets, datasets and models by name, and the values stay on the server. Over MCP it works with a personal token that has your permissions in one workspace and cannot manage tokens, workspaces or your account.
No. Choosing who labels or reviews is left to you. The agent leaves those inputs empty and tells you which ones need a person before the flow can run.
Claude Code, Codex and any client that adds a Streamable HTTP MCP server with a request header. Clients that only sign in with OAuth, such as the connector screen in Claude Desktop, cannot connect yet.
Yes. The administrator adds an OpenAI or Anthropic API key for AI Chat, and the MCP server runs at your own OvalBee address followed by /api/mcp.