Describe the goal. Let it compose the steps.
Share your inputs, expected results, and business rules. Workrun plans how to get there using the Agents, Apps, and tools in your workspace.
CONVERSATION → WORKFLOW · LOCAL-FIRST · REUSABLE
Tell Workrun what you want to achieve. It works out the steps, selects the capabilities, and turns your plan into a reusable workflow. Refine it through conversation or open the canvas to edit the details. When a matching workflow already exists, ask Workrun to find and run it.


Start with a conversation. Keep refining it in chat or open the generated workflow on the canvas.
Describe the goal. Let it compose the steps.
Share your inputs, expected results, and business rules. Workrun plans how to get there using the Agents, Apps, and tools in your workspace.
Shape the workflow as you talk
Add conditions, adjust steps, and refine the plan together. Approve it to create a workflow, then keep editing through conversation or on the canvas.
Build once. Use it again.
Your workflow stays in your workspace. Bring new inputs, review the selection, and run it again, with each step and result available to inspect.
Learn how to build workflows through conversation, or start with Apps and the canvas below.
APP · START WITH CODE
Every App is a local uv project with its own dependencies,
entry-point code, and data contract. Run it on its own for file work,
internal integrations, or human input; a workflow is optional.
Later, use it as a Process node or an Agent tool when it belongs in a flow
Build your first Python App
WORKFLOW · START WITH COLLABORATION
Connect capabilities when a task needs judgment, branches, and people.
Compose Agents, Process Apps, conditions, subworkflows, and human review on a canvas. Node state, tool traces, and results stay visible, so complex automation remains inspectable and resumable.
Define what each step reads, produces, and hands to a decision maker
Iterate safely with evaluations, version comparison, and quality gates
Build your first workflow
For data processing, file operations, internal integrations, or a local tool that needs input from a person, an App can run directly from Apps. Workrun creates and runs the project, preserves run history, and handles runtime interactions; the Python remains your codebase.
Run on demand or on a schedule
Run a local project directly from Apps, or schedule recurring collection, cleanup, and synchronization work.
Let code control runtime interaction
Python can request a JSON-Schema form or confirmation. After a person submits, rejects, or cancels, the code decides what happens next.
Join a flow when needed
The same App can become a fixed-position Process node; a Tool App lets an Agent call it on demand through a schema.
Open the canvas to inspect and refine the workflow created through conversation. Connect Agents, Apps, and branches, and check how data moves between steps. After running it, inspect node messages, tool calls, and results.

This order-cancellation workflow checks an order, routes by risk, and hands high-risk cases to a separate branch. See how each step connects before you run it.
Tool authorization, question responses, and human review appear directly in node messages, so you can inspect output and handle requests later. See the screenshots and instructions.
The video starts with “Plan a trip based on the weather in Hangzhou.” After choosing to begin, the reviewer changes “Hangzhou” to “Shanghai” in input, then approves and continues. A weather Agent calls its tool, and a trip-planning Agent reads the result to suggest an itinerary. Questions, human review, and tool authorization all happen in node messages.
Hangzhou trip request → Choose to begin → Edit destination to Shanghai → Authorize weather lookup → Get Shanghai weather → Suggest a Shanghai itineraryTry this example: Download the example Workflow (ZIP), including the workflow and weather App source code. The example’s prompts and App name are in Chinese.
See the workflow guide for import and export details.
Review can change the actual content. The original request is “Plan a trip based on the weather in Hangzhou.” This review node uses input as its content key with Allow editing enabled; the reviewer changes “Hangzhou” to “Shanghai,” then approves and continues. Approval writes the edit back to that same state key, so the tool queries Shanghai weather and the planner follows the revised request. Check city: "上海" (Shanghai) on the tool authorization card to see that the edit reached downstream execution.
| Step | Who handles it | What it produces |
|---|---|---|
| Confirm requirements | You select, edit, or approve on question and review cards | Confirmed trip requirements; choosing not to plan or rejecting review ends the run |
| Check the weather | The weather Agent calls its configured tool | Weather information in weather_info, accessible to the trip-planning Agent |
| Plan the trip | The trip-planning Agent reads the requirements and upstream weather | Itinerary suggestions grounded in the tool result, without inventing weather |
The video demonstrates orchestration and interaction. Actual weather depends on the connected tool’s data source and query time. Suggestions based only on current weather are a reference, not a forecast for a future trip. Learn about configuring human interactions and passing state between nodes.
| What you need to do | How Workrun helps |
|---|---|
| Compose capabilities | Run an App independently, or connect Agents, Processes, branches, subworkflows, and human nodes on a canvas. |
| Run reliably | Isolated node state, input-pause resume, interruption recovery, and fresh reruns after failure. |
| See what happened | Inspect model messages, tool calls, script logs, and traces node by node. |
| Iterate safely | Validate workflow changes with evaluation suites and quality gates before publishing. |