GPUI Agentation

io.github.emmanuel-defreitasPublished by GitHub account emmanuel-defreitasv0.6.0Updated Nov 2, 2026

Read and manage development annotations from a local GPUI app.

Overview

AI-generated overview

Loopback MCP server from a GPUI debug app that lets a coding agent read and manage UI annotations with element bounds and Rust source locations.

What you can do
  • Read annotation sessions and notes captured in a running GPUI debug app, including element ID, label, bounds and Rust call site
  • Export feedback as Markdown for pasting into an agent conversation
  • Work through requested feedback and update note status (pending, acknowledged, resolved, dismissed) and reply threads
  • Discover the connected MCP tools and inspect registered screens and finer targets
When to use it
Useful when you are developing a native GPUI (Rust) app and would rather point at a UI element than describe a bug in words or paste a screenshot, and want your coding agent to read those notes directly.
Requirements
The crate must be added to your own GPUI app and initialized in debug builds only; the debug app must be running with the MCP server switched on in Settings → MCP. Requires macOS with Xcode command-line tools and Rust via rustup. The endpoint runs on loopback; clients that need stdio use a bridge. No authentication or environment variables are declared.
Before you install
It is a development tool: release builds expose no annotation API, and it annotates registered targets rather than arbitrary pixels. Notes persist to a local JSON file, and agents can change note status and reply threads, so writes should be reviewed. The loopback endpoint has no declared authentication.

Installation

In SourceWeft

  1. Open GPUI Agentation in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.

Other MCP clients

Follow the launch instructions in the repository.

README

[Contributors] [Forks] [Stargazers] [Issues] [MIT License]


[gpui-agentation logo]

gpui-agentation

🎯 Point at a running GPUI app, leave a note, and your coding agent gets the element, its bounds, and its Rust source location.
Explore the docs »

View Demo · Report Bug · Request Feature

Version: 0.6.0. Debug builds only — release builds expose no annotation API. Pinned to gpui-pre =0.3.8.

Table of Contents
  1. About The Project
  2. Getting Started
  3. Usage
  4. Try the Demo
  5. Roadmap
  6. Contributing
  7. License
  8. Contact
  9. Acknowledgments

About The Project

[gpui-agentation dashboard demo]

Agentation annotates web pages through the DOM. A native GPUI window has no DOM, so gpui-agentation registers screens (and optional finer targets) with an ID, label, bounds, and Rust call site, then attaches your note to that.

Use it when you would otherwise describe a UI bug to your agent in words or paste a screenshot. Notes persist to a local JSON file, carry a status (pending, acknowledged, resolved, dismissed) and a reply thread, and an agent reads them over a loopback MCP server or REST/SSE.

It is a development tool. It annotates registered targets, not arbitrary pixels (by default, the whole screen), and release builds do not include it. For production telemetry or session replay, use something else.

(back to top)

Built With

[Rust] [GPUI] [MCP]

(back to top)

Getting Started

Add gpui-agentation to your own GPUI app in three steps: install the crate, initialize once at launch, then mount and wrap your screen.

Prerequisites

  • macOS with Xcode command-line tools
  • Rust via rustup (the repository pins the toolchain)
  • Python 3 (only needed to run the bundled demo)

Ely is bundled, so there is nothing else to install.

Installation

  1. Add the crate with the native feature set:

    sh
    cargo add gpui-agentation --no-default-features --features native
  2. The crate does not exist in release builds, so every call sits behind #[cfg(debug_assertions)] and release returns your content unchanged:

    rust
    // Launch: serve Agentation's assets, then initialize.let app = gpui_platform::application();#[cfg(debug_assertions)]let app = app.with_assets(gpui_agentation::Assets);app.run(|cx| {    #[cfg(debug_assertions)]    gpui_agentation::init(cx).expect("initialize Agentation");    // ...open your window});
    // Window setup: app name, route, notes file, local port. Keep `feedback`// in a field of your root view that carries the same cfg.#[cfg(debug_assertions)]let feedback = gpui_agentation::Agentation::mount(    "My app", "gpui://my-app/home", "target/agentation/feedback.json", 4748, window, cx,)?;
    // Render: wrap the screen in debug builds; release keeps `content` as is.#[cfg(debug_assertions)]let content = self.feedback.root("home", "Home screen", content, window, cx);content
  3. Add target(id, label, child) only where you want finer inspection. Ordinary controls need no wrappers. Agentation handles its Ely setup, assets, and focus internally.

The consumer example keeps all of this in one devtools module. More: setup guide, setup skill for agents, and customization.

(back to top)

Usage

Inspect and annotate

Turn on the power switch, then choose Inspect, Box, Arrow, Pen, Comment, or Textbox. Select an element or draw on the screen. Inspect surfaces the element's ID, label, bounds, and Rust call site.

[Inspecting a registered element]

Draw an arrow (or a box or freehand pen stroke) to point your agent at exactly the right spot.

[Drawing an annotation arrow]

While annotation mode is on, use I / B / A / P / C / T to select a tool. Hover a tool to see its shortcut as an Ely keycap in the tooltip. Ctrl/Cmd+S saves the current draft; Ctrl/Cmd+Z and Ctrl/Cmd+Shift+Z undo and redo annotation changes. Delete/Backspace removes a selected annotation when its text field is not being edited. Text fields retain their ordinary typing and text undo behavior.

Textboxes use plain text with one style and size, preceded by a numbered pin. Click to select and show the bounding box; drag the text or pin to move both. Double-click or press Enter on the selected textbox to edit it. Saving displays the text directly on the canvas without opening a comment. Other numbered pins can be dragged independently of their drawings.

In a floating comment, type @ to choose an existing note or #12 to preview pin 12. Hovering a suggestion or using the arrow keys animates its pin; click, Enter or Tab inserts an @#12 reference. Escape dismisses the list. Right-click an annotation for Edit, Delete and Color actions, with undo support.

Include gpui_agentation::debug_menu() in your host's cx.set_menus(...) list to expose the controls in the native Debug menu. Agentation::root routes those menu commands to the active window. Settings → Toolbar → None hides the embedded toolbar. Toolbar placement persists per application in its notes file and is shared by windows mounted against that file.

Capture feedback

Write a comment and save it. Open Comments → Copy to copy your feedback as Markdown, ready to paste to your agent.

[Copying comments as Markdown]

Move the toolbar out of the way whenever it overlaps the UI you are inspecting.

[Repositioning the toolbar]

Connect your agent

  1. Keep the debug app running.
  2. Open Settings → MCP and turn the server on.
  3. Copy the config into your MCP client.

The endpoint runs on loopback. Clients that require stdio can use the MCP bridge.

  • MCP Registry name: mcp-name: io.github.emmanuel-defreitas/gpui-agentation

For more examples, please refer to the Reference.

Agent skills

Give your coding agent a reusable workflow for installation or exploring a running app's feedback. Both skills live under skills/:

SkillUse it to
gpui-agentation-setupIntegrate the native library into a GPUI app, register screens, and verify debug and release builds.
gpui-agentation-exploreDiscover the connected MCP tools, read sessions and notes, export Markdown, and work through requested feedback with verified status updates.

From a clone of this repository, copy either skill directory into your host project's .agents/skills/ directory, or into ~/.agents/skills/ for personal use across projects. For example, to install the exploration skill in a host:

sh
mkdir -p /path/to/your-app/.agents/skillscp -R skills/gpui-agentation-explore /path/to/your-app/.agents/skills/

In Zed, start a new conversation after copying, then invoke /gpui-agentation-explore or /gpui-agentation-setup. Other agents can load the linked SKILL.md directly or use their supported skill directory. Try: “Use gpui-agentation-explore to discover my app's MCP tools and summarize pending feedback without changing notes.” Skills provide instructions; connect the running debug app's MCP server separately using the steps above.

(back to top)

Try the Demo

Clone the repository and run the bundled demo app:

sh
git clone https://github.com/emmanuel-defreitas/gpui-agentationcd gpui-agentationmake dev
  1. Choose Inspect, Box, Arrow, or Pen.
  2. Select an element or draw, then write a comment and save.
  3. Open Comments → Copy to copy your feedback as Markdown.

For rebuilds on save, enable Settings → Hot reload → Server status. It rebuilds and restarts the app after a successful build; in-memory state resets, saved annotations survive. Demo details

(back to top)

Roadmap

  • Inspect, Box, Arrow, and Pen annotation tools
  • Loopback MCP server plus REST/SSE access
  • Hot reload for rebuild-on-save
  • Add text annotation tool
  • Update Comment annotation UX
  • Add gpui-devtools features from joris-gallot/gpui-devtools
  • Extends AI features from hexuria/gpui-agent
  • Stabilize the API toward 1.0
  • Port MCP tools to a standalone macOS companion app

See the open issues for a full list of proposed features (and known issues).

(back to top)

Contributing

Contributions are what make the open source community such an amazing place to learn, inspire, and create. Any contributions you make are greatly appreciated.

If you have a suggestion that would make this better, please fork the repo and create a pull request. You can also simply open an issue with the tag "enhancement". Don't forget to give the project a star! Thanks again!

  1. Fork the Project
  2. Create your Feature Branch (git checkout -b feature/123-amazing-feature)
  3. Commit your Changes (git commit -m 'feat: add amazing feature')
  4. Push to the Branch (git push origin feature/123-amazing-feature)
  5. Open a Pull Request

Changes target dev, then promote through next to main. Run make ci for checks or make help for available commands.

Before promotion, run make release-prepare to calculate the next stable version from conventional commits and synchronize Cargo, the lockfile, Registry metadata, documentation, skills, and site references. Settings, CLI, and MCP versions derive from Cargo at build time. Review that diff and update CHANGELOG.md before committing through the promotion PRs. CI and release tagging reject mismatched versions. See the release policy.

(back to top)

License

Distributed under the MIT License. See LICENSE for more information.

(back to top)

Contact

Emmanuel De Freitas — @emmanuel-defreitas

Project Link: https://github.com/emmanuel-defreitas/gpui-agentation

(back to top)

Acknowledgments

(back to top)

Source: README.md at commit b0a8442

Tools

0
SourceWeft started this package once in an isolated sandbox, without credentials, to list its tools. That failed: the package type cannot run in the sandbox.

Version history

5
  1. v0.6.0LatestOct 26, 2026
  2. v0.5.1Oct 23, 2026
  3. v0.4.0Oct 21, 2026
  4. v0.3.0Oct 21, 2026
  5. v0.2.0Oct 20, 2026

Score

50% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 days25 pts

    Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.

  • Auth declared10 pts

    Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.

  • Tools listed15 pts

    Answer tools/list without a credential, or ship a package that starts in a sandbox without one.

  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts