Getting started with Studio

Open Playground from the website navigation to try Studio without a service. It includes a complete research chain and two backtests. In task submission, select playground.backtest, choose strategy and outcome, then follow progress, logs and results in the task center or cancel execution. Successful tasks appear in Backtest; Strategy comparison compares the sample runs. Agent replies are scripted. All data and execution are simulated; refresh or use Reset demo to restore initial state.

AxonX Studio is a workbench for tasks and research. It reads Job and Task definitions provided by the service, generates submission forms from Schemas, and displays run status, logs, dependencies, and standard research artifacts.

The screenshots on this page use the English interface; the documentation is available in English and Chinese. Screenshots show actual pages, and the visible task and plugin counts depend on the current machine. Task identifiers, private paths, remote addresses, and browser tabs from real experiments are excluded from documentation images; docs-demo and docs-child in task-operation screenshots are built-in demos created specifically for the documentation.

Open Studio

Install AxonX with Studio, then start the service (axonx[full] also includes Studio):

pip install "axonx[studio]"
axonx start --service.host 127.0.0.1

Configure the Service token before starting, then open http://127.0.0.1:1024/. AxonX loads static assets from the axonx_studio Python package. The API still works when Studio is not installed.

For frontend development and proxy configuration, see Studio development.

Configure the connection

Open Settings in the upper-right corner and enter the token used by the local service in Service token. The browser saves the local token; use the same settings entry to clear it. The screenshot only shows “Token configured” and contains no credential value.

The machine selector defaults to Local. Once remote targets are configured, the selector lists the service's targets. Studio continues to send requests to the same-origin backend; the remote address is passed through the outer request's target, and remote tokens are stored in backend configuration.

Switching machines changes the available task definitions, run records, and research results. Confirm the current target before operations, especially submission, cancellation, and deletion. See Using remote machines for connection differences.

From submission form to task details

1. Open Submit task and select demo under Native tasks.
2. Enter X and Y, for example 2 and 3; leave Fail as False.
3. Optionally enter a Task Name for easy lookup, or leave it blank to generate a name. Source Tasks accepts Task IDs separated by ASCII commas.
4. Click Submit run, then open Task management after a successful submission to view the run.
5. Open task details and check the state, steps, logs, dependency graph, and final output.

The form comes from the Task's input_schema. Required fields, enumerations, booleans, numbers, and JSON objects are handled according to field types; blank optional fields are not automatically submitted as empty strings. The backend still performs actual model validation, and complex Schema constraints cannot rely solely on form rendering.

Rerunning a fixed instance name replaces a completed task directory. To retain multiple experiments, use different Task Names or generated names. See Lifecycle.

Use task and resource pages

Both demos in the screenshot succeeded; see Task management and Task lineage for operation details, logs, and relationship graphs.

Task management provides filters for task type, state, and name, and supports automatic refresh, details, cancellation, and multi-selection deletion. Active task details show progress and logs through an event stream; completed logs can be read on demand.

The dependency graph shows tasks and upstream relationships. Relationships for running tasks may come from status configuration, while successful task relationships come from metadata; missing upstream records appear as missing nodes. Graph relationships are not an automatic execution plan.

Machine resources shows CPU, memory, and GPU metrics for the current target. Device data depends on the machine and available vendor tools; missing GPU data does not necessarily indicate GPU failure, nor does it mean the framework automatically selects machines or allocates devices.

View research results

Page	Main content	Details
Tushare data	Raw data directories and Parquet previews	Data download
ETL	Row counts, date ranges, features, and labels	Interpreting results
Factor analysis	Factor scores and metric groups	Interpreting results
Model training	Model configuration, metrics, and training curves	Artifact protocol
Offline prediction	Prediction data, statistics, and artifacts	Interpreting results
Offline backtest	Daily curves, quality, and period summaries	Interpreting backtests
Strategy comparison	Comparison of two backtests over their shared interval	Strategy comparison

Research pages read metadata.json and output_params.artifacts. A run record in the task list does not guarantee that displayable research metadata has been produced. For failed tasks, incomplete fields, or corrupt metadata, investigate details and logs first.

Deletion on research pages removes the selected directories and artifacts; downstream dependencies are not automatically rebuilt. Confirm which experiments must be retained before deleting; see Workspace files for operation semantics.

Agent and API entry points

Agent supports resuming sessions, viewing text and tool calls, stopping the current turn, and renaming, tagging, forking, and deleting sessions. The explanation entry in task details can pass the Task ID to Agent, but an actual conversation requires available model configuration. Stopping a conversation does not cancel a Task.

API interfaces generates call forms from the current target's Job catalog. It calls Jobs rather than directly executing arbitrary Tasks: for example, select get_task_definition to inspect a definition, then use submit to submit a registered task. Ordinary calls return responses; see SSE events for streaming protocols.

The screenshot shows an actual call to the read-only version interface, returning HTTP 200 and a version string. Other interfaces may submit tasks or change data; check their semantics before calling them.

Page links and preferences

Page routes use #machine/page/view/resource, for example #local/task-defs/catalog/demo. The router encodes the resource portion; when copying a Task ID to the command line, quote it as a complete string.

Language, theme, and panel width are saved in browser preferences and do not change server-side research configuration. All screenshots on this page use English; switch to Chinese through the language button at the top.

Feature screenshot index

Feature	Documentation containing screenshots
Home, demo submission, settings, API debugging	Quick start and this page
Task list, details, and logs	Task management
Parameter snapshots and relationship graphs	Task lineage
CPU, memory, and GPU	Remote machines
New Agent session	Agent usage
Data download form and Parquet preview	Tushare data
ETL, factors, training parameters and curves, prediction	Research results
Returns, quality, positions, overall and annual/quarterly/monthly summaries	Interpreting backtests
Five strategy comparison tabs	Strategy comparison
Notification task form	Notifications

Empty state on first use

Before any tasks have run, the list shows an empty state. Submit this page's demo to produce the successful records shown earlier; a complete research workflow is not required to verify the connection.

When problems occur

Symptom	Check first
Page opens, but API returns 401	Whether the local token matches service configuration
Task definitions or API catalog are empty	Whether the service has a token, plugins are installed on the selected machine, and Jobs are public
Research page has no results	Whether the corresponding metadata was generated and contains standard outputs
Charts lack curves	Whether fields or files such as training_curve, daily, and summary exist
Remote machine request fails	Backend targets addresses, tokens, and target service connectivity

See FAQ and Troubleshooting and recovery for more information.