Use case · AI agents
AI agents you can inspect
An agent is a loop: prompt, tool call, result, next prompt. On the canvas each turn is a node or a group, and the console shows exactly what went in and what came back, so a bad answer is a graph to read rather than a black box.
What you build
Every graph is a set of small Python nodes wired by ports. These are the pieces this track puts together.
Prompt steps
The AI category ships the LM Studio Chat node, which talks to any OpenAI-compatible endpoint and keeps a per-chat history. Closed prompt nodes from the store expose only their manifest and ports.
Tool calls
Pair a model node with HTTP Get or HTTP Post to call your own tools, then use a Branch node to choose the next step from the reply.
Memory and state
Keep conversation state in node parameters or in a memory node. Because the graph is visible, you can show a reviewer where the context is added and where it is dropped.
Build it step by step
The same five moves work for every graph: add nodes, wire ports, set parameters, run, then save or share.
Open the workspace
AI Agent Workspace is a public project with one Agent graph. Open it in the editor to see the starting shape before adding anything.
startprintAdd a prompt node
Use LM Studio Chat from the catalog for a local model, or rent the closed OpenAI Prompt Node from the store when you only need the interface and not the source.
lmstudio_chatWire one turn
Send the user message into the prompt node, take its reply to a Print or a send node, and keep the exec chain continuous so each step runs in order.
printGive the agent a tool
Add an HTTP Get node, call your API, and feed the response into the next prompt. A Branch node can choose between answering and calling another tool.
http_getbranchRun and inspect each turn
The console lists every node event with its output. If a prompt is wrong, you can see whether the model, the input or the tool response caused it.
Save, share or sell
Save the graph, make the project public for readers, or snapshot it into the store as an open or closed asset with a free, one-time or rental variant.
Example you can open now
This example is seeded on the public site. Open the graph in the editor to inspect it, or take a copy through the store when a fork is not enough.
AI Agent Workspace
An open project with one agent graph and a clean structure to extend. The store variant adds the source download, so you can keep a copy outside the public project.
How it runs
The API starts a container for the graph and forwards the run to the orchestrator. Prompt and tool nodes are ordinary nodes: they read their inputs, await their work and push results to their output ports. Logs and node events stream back while the loop is running, so a slow model call is visible as a running node rather than a frozen page.
- Each graph runs in its own Docker container, started by the API.
- Nodes are fetched at the exact version the graph pins.
- stdout and node events stream to the editor console while it runs.
Build your own agent loop
Start with one prompt node and one tool, then split the loop into a group once it works. Small graphs are easier to debug and cheaper to run.