Typed operations
Clear inputs and outputs make actions easier to inspect and reuse.
Give agents an interfaceto the computer.
Discover capabilities, call typed operations, and keep execution behind explicit permissions and one shared policy boundary.
Install SemwrightPublic installer arrives with the verified release.
release installer not published yetA practical bridge between agents and your tools.
Semwright gives agents a common way to discover what software can do, address the right object, and request an action without turning every application into a separate automation project.
Native integrations
Use application APIs and semantic interfaces where they exist.
You stay in control
Permissions and policy remain separate from the requesting agent.
Works with the tools you rely on.
Different applications expose different interfaces. Semwright keeps the execution model consistent while each driver stays explicit about its own scope and limits.
Chromium
Work with tabs and the DOM, not an estimate of where a button might be.
ExperimentalLibreOffice
Read and write documents through a deliberately narrow UNO interface.
ExperimentalBlender
Reach into the scene through an application-native driver.
ExperimentalKiCad
A bounded PCB interface for an engineering tool with its own object model.
ExperimentalMLT video
Address timelines, tracks and clips as structured objects.
ExperimentalLinux desktop
Find semantic controls and use native window interfaces where available.
ExperimentalYour agent gets access. You keep the keys.
Semwright separates what an agent can discover from what it is allowed to do. Input, clipboard access, capture and app changes stay behind explicit policy instead of riding along with the frontend.
See how permissions workChoose what an agent may do.
Discovery can stay broad while sensitive actions remain separately controlled.
InputKeyboard and pointerGrant required
ClipboardRead or write clipboard dataGrant required
CaptureScreen and screenshot accessGrant required
App changesNative write or destructive operationsPolicy checked
Point to meaning, not pixels.
Semwright returns semantic references that identify the object an agent actually found.
- Application
- org.gnome.TextEditor
- Object
- Save
- Role
- button
- Reference
ui:window/toolbar/save
If a reference is stale or ambiguous, the request stops instead of guessing.
One route through the computer.
CLI, MCP and recipes reach application providers through the same broker boundary.
FrontendCLI, MCP, recipe
SemwrightPolicy and references
ProviderDriver or native API
Changing the frontend does not create a more privileged path.
One request. The right tool. A controlled action.
Your agent asks Semwright for a capability. Semwright resolves the application object, checks the current reference and authority, then dispatches through the matching provider.
Follow a real requestSave a documentHow Semwright handles one typed request
Shared execution pathAsk for the thing you mean.
The agent calls a typed operation instead of guessing a screen coordinate.
- Operation
ui.find- Application
- Text Editor
- Looking for
- Save button
Get a semantic reference back.
Semwright identifies the object and reports the actions that object exposes.
Matched objectSave
ui:window/toolbar/saverole: button / action: clickPermission stays in the path.
The request can only dispatch when the current reference and required authority are both valid.
- Operation
ui.invoke- Action
- click
- Dispatch
- Text Editor provider
No pixel guessing.No frontend-specific bypass.
Questions before you connect an agent.
Installation, permissions, MCP and platform scope, explained without hiding the current limits.
Is Semwright another AI agent?
No. Semwright is the interface an agent calls, not the model that decides what to do. It exposes capabilities and mediates execution. You bring the agent.
Why not just give an agent a shell?
A shell usually inherits broad user authority. Semwright is built around narrower operations, scoped references, explicit permissions and one audit path.
Is this only an MCP server?
MCP is one frontend. The CLI, inspector and recipes use the same broker. Application drivers and federated MCP providers extend the capability catalogue.
What can I use today?
The development source is available for Linux. Application and desktop coverage is experimental and version-specific; the compatibility guide separates code, test evidence and live availability.
Will Windows and macOS be supported?
Cross-platform support is part of the product direction. Linux is the first host. This site does not present Windows or macOS as currently supported.
The computer has more to say. Give your agent a way in.
Explore the architecture, try the source, and inspect the boundaries.