Build with an AI agent
An agent in OvalBee edits the same flow you see on the canvas. It looks up nodes in the Node Library, reads the datasets, models and secrets available in your workspace, writes the graph and binds inputs — and every change shows up on the canvas as it happens. You can keep editing by hand at any moment.
There are two ways to work with one:
What an agent can do
Both kinds of agent use the same set of tools, so they can do the same things:
- Build and change flows — add, remove and rewire nodes, set inputs, bind references between nodes, set schedules and add sticky notes.
- Find the right node — search the Node Library by task, read a node’s inputs and outputs, and look up the valid choices for an input (for example, which datasets or pretrained checkpoints exist).
- Explain — read a flow and tell you what it does, or why a node failed by looking at its recent runs.
- Undo its own work — the agent can list its earlier edits and roll one back. On top of that, on plans with flow version history the flow editor saves a version named Before agent edit before the agent’s changes are applied, so you can always restore the flow from Versions yourself.
- Run and stop — start a node, or stop what is running, if your role allows running flows.
Every write is validated before it’s saved: a node that doesn’t exist, an input value that isn’t allowed or a reference to a missing output is rejected with an explanation, and the agent corrects itself. Nodes the agent doesn’t touch keep their settings and run history. If someone changes the flow while the agent is working, its write is refused and it re-reads the flow first, so it can’t overwrite your edits.
Agents never see secret values. They refer to secrets and datasets by name, and the actual credentials stay on the server.
AI Chat in the editor
A chat is yours within a workspace, not tied to one flow, and it follows you around the app. Open it from AI chat at the bottom of the sidebar, or from AI Chat in the toolbar above the canvas. On an empty flow the panel opens by itself.
Outside the editor the chat is for questions about the workspace — which flows exist, which of them failed recently — and for starting a new flow. Ask it to build one and it creates the flow and links to it; open the link and the same conversation moves next to the canvas, where “this flow” now means the one you’re looking at. The agent creates a flow only when you ask for one, so asking questions leaves nothing behind.
Type what you need, or start from one of the suggestions:
- “Build a flow that imports my images, splits them 80/20 and trains a YOLO detector”
- “Explain what this flow does”
- “Find and fix missing inputs”
- “Add an evaluation branch that compares the new model with the current one”
The model
An instance administrator picks the model and sets the API key of its provider under Admin → Configuration → Integrations. On OvalBee Cloud every answer counts against your plan’s monthly credits, shown by the ring next to the space selector. A chat keeps the model it started on: when the administrator switches to another one, older chats stay readable and the panel asks you to start a new one.
The agent asks a question when a choice is really yours — which dataset, which classes, how big a validation split — instead of guessing. Before it runs a node, stops a run or reverts the flow to an earlier revision, it shows the action and waits for you to approve it. Your conversations are listed in the panel header, the same list on every page; switch between them, start a new one or delete the one you have open. A deleted chat can’t be brought back, and its messages are erased from the installation 30 days later, or after the period the administrator set. A new chat is saved once you send its first message. Opening a flow resumes the chat you were having, or else the latest one started in that flow.
Until an instance administrator sets the key, the panel says the agent isn’t configured. Nodes that call a language model, such as Dataset insights, run on the same key and the same credits, and their run is refused when the key is missing or the credits are used up.
Connect your own agent over MCP
OvalBee runs a Model Context Protocol server, so an agent you already use can drive OvalBee directly. It has the same tools as AI Chat; choose it when you’d rather work in your own client, for longer sessions, or alongside your own code and data.
Open the connection dialog
On the Home page, click Build with your agent. The dialog shows the server address, which is your OvalBee address followed by /api/mcp.
Generate a token
Click Generate token. The token is shown once — copy it now. It’s listed afterwards under Settings → API tokens, where you can revoke it.
Claude Code
Codex
Other clients
Run this once, in any directory. Use single quotes around the header — the token contains characters a shell would otherwise expand.
Then start a new Claude Code session. A session that was already open won’t see the server.
In a client that supports MCP Apps, the flow is drawn right in the conversation whenever the agent reads, changes or creates one. It looks like the canvas: the same nodes, icons and connections, with the nodes the agent just added or changed outlined. You can drag the graph, zoom, and select a node to see what it is wired to. Run flow starts the flow from its first node, and Run from here starts it from the node you selected; both ask before anything runs, and each node then shows whether it is waiting, running, completed or failed. While something is running, Stop cancels it. Other clients get the same answers as text.
The token acts as you: the agent sees the same workspace and can do exactly what your role allows. A token belongs to one workspace, so create one per workspace if you work in several. Creating a new token from the dialog replaces the previous agent token, which stops working.
Getting good results
- Name the real things. “Train on the
PCB componentsdataset, classesresistorandcapacitor” beats “train on my data”. The agent can look up names, but it can’t read your mind. - Say what should happen to the results. Where accepted images should end up, which model to compare against, what threshold counts as good.
- Ask it to explain before it changes. “What would you change to add a review step?” gets you a plan to agree on first.
- Keep runs deliberate. Building a flow never runs anything. Training and prediction nodes use real compute, so the agent asks before it starts one — and you can always start runs yourself from the canvas.
- Iterate on the canvas. Once the structure is right, small changes such as a threshold or a batch size are faster to make in the node’s sidebar yourself.
Related pages
- Flow editor — what you can do on the canvas by hand.
- References — how nodes pass data, which is what the agent sets up for you.
- Secrets and API tokens — managing the token your agent uses.