A bounded workspace
List, read, search and inspect files within admitted roots. Tool arguments do not expand access: the host checks the operation and its target before execution.
- File access
- Permission checks
Pyrodit Stack / Local agent execution
Pyrodit Runner
Connect agents to files, commands, browsers and selected applications through explicit permissions. Follow each operation from request to recorded outcome.
Product interface · Sample workspace · English UI
Your desktop. A clear operating boundary.
Start a task, choose the model and review application and folder access in the Runner workspace.
Folder access beside the task
Review the selected access mode and folder scope directly from the conversation composer. The visible choice describes this workspace; individual tool calls remain subject to Runner policy.
Choose how the agent connects
The settings bring API and local Ollama models, foreground or background application focus and the tool-access summary together. The selected model and operating mode remain explicit.
What you can do
- 01
Work inside an explicitly authorised workspace.
- 02
Separate observation from permission to act.
- 03
Inspect operation records, outputs and uncertain outcomes.
Inside the product
Pyrodit Runner
Commands with a visible lifecycle
Prepare a command, evaluate policy, start it and inspect status or bounded output. Cancellation and process records keep long-running work visible beyond a single UI interaction.
- Command execution
- Task monitoring
Browser work with explicit targets
Navigate, observe and interact with browser pages through typed tools. Observation and input have separate permission checks; operations remain bound to their browser session and recorded evidence.
- Browser interaction
- Session boundaries
Selected applications, specific permissions
Discover and open installed applications, observe a selected window and use supported window-bound actions. Screen capture, accessibility inspection and input are distinct capabilities; a screenshot grant is not permission to control the desktop.
- Application control
- Device permissions
Evidence beyond the agent’s answer
Keep operation receipts and image artifacts separately from chat text. Unknown outcomes remain explicit; reconnecting does not silently replay a potentially completed action.
- Activity history
- Result tracking
One workspace for operators and integrations
Review chat, process records, tools, workspace settings and OS permissions in the desktop interface. SDK-backed tool contracts provide an integration boundary while planning and workflow orchestration remain separate responsibilities.
- Desktop workspace
- SDK integration
Scope & availability
Runner supplies the local execution layer; Agentic Runtime coordinates the workflow. OS permissions and enabled host capabilities define access. Native application control depends on platform and integration; browser sessions use an explicitly owned target. Screen images can contain sensitive information and are shared with the configured model when vision is used.