A new interface for working with agents
RambleDesk wants to move collaboration beyond the chat box.
RambleDesk is a desktop workbench for working with coding agents. When you need to try their work, point out a problem, or decide what happens next, it gives you a place to do that.
Leave suggestions beside a document, mark a problem in an image, or write or speak your thoughts. RambleDesk sends your feedback together with the content it refers to, so the agent can use it to continue the task.
Download · Get started · Documentation
Agents can write code, research a topic, and run tools on their own. But moving a task forward still often needs your involvement:
- Add context they don't have. Explain the business background, clarify a requirement, or answer a question.
- Choose a direction. Compare approaches, express a preference, or set priorities.
- Review what they wrote. Read a plan, a draft, or code, and suggest changes in the relevant places.
- Try it yourself. Open a page, use an interface, or run a CLI, and note what works and where you get stuck.
- Check the result. Verify table values, details in an image, or a passage in audio or video.
If you've used askUserQuestion, you already know one kind of workbench interaction: the agent presents a question and options, asks you to answer, and continues based on your response. RambleDesk's questions workbench follows that familiar pattern and extends it to other tasks. Reviewing a document brings up the text and annotation controls; trying a page brings up the page and a place for feedback.
A workbench prepares the material and controls for a task that needs your input, then collects your response for the agent. You don't have to turn every judgment into a chat message. Answers, annotations, and attachments return to the agent with your feedback. In sessions started inside RambleDesk, submitting feedback queues the next turn in the original conversation.
RambleDesk provides eleven workbenches in the current source:
| Workbench | What you can do |
|---|---|
| Questions | Answer questions one by one, choose options, and add explanations, much like askUserQuestion |
| Free feedback | Follow review steps and record feedback with text, voice, and attachments |
| Document review | Review Markdown and annotate selected text, cross-paragraph ranges, or the whole document |
| Image annotation and canvas | Draw arrows, boxes, and text on an image or blank canvas to show where and what to change |
| Web review | View a page and leave comments on specific elements |
| Terminal trials | Use a real CLI, quote its output, and record your experience |
| Drag sorting | Reorder options to express priorities or preferences |
| Brainstorming | Develop idea cards, group and select directions, and carry your edits into the next round |
| Diff review | Review code changes and comment on lines, ranges, or hunks |
| Table review | Suggest cell values and add comments |
| Audio and video review | Play the material and leave feedback on a timestamp or segment |
Annotations and proposed rewrites are saved as feedback and suggestions. Actions in a web page or terminal have their usual effects. See the workbench playground for more examples.
See a few ways to work with an agent.
Select a passage, add a comment or suggested rewrite beside it, then keep reading.
Mark an area and explain the adjustment so the agent knows what you're referring to.
Edit and group idea cards, select directions, then ask the agent to expand, combine, or continue with your choices. Your edits carry into the next round.
Follow the agent's steps to try the result. Write or speak your observations, and attach supporting material when needed.
You can also add text, voice, and attachments. Optional AI cleanup helps organize your wording while preserving your original feedback. Unfinished feedback is saved as a draft.
Download 0.5.0 for Windows x64 or macOS Apple Silicon from GitHub Releases.
- Connect an agent. Open Settings → Agents. Follow the prompts to install supported components or connect an existing agent such as Claude Code, Codex CLI, or Gemini CLI. Sign-in and model access belong to the agent itself.
- Give it a task. Start a session, choose an agent and project folder, and describe your goal.
- Try it and respond. When a feedback request arrives, review the result, follow its steps, and submit your comments. Feedback returns to the original conversation so the task can continue.
Open Settings → External adapters and follow the integration guide. How the task continues after delivery depends on the adapter; Generic MCP requires returning to the agent manually.
Text feedback needs no extra setup. For speech, download a local transcription model in Settings → Voice and allow microphone access. AI cleanup is optional and needs a separately configured model service.
The Windows installer is not Authenticode-signed; macOS builds are not notarized. See the installation guide.
Before upgrading from 0.4.0, make a complete backup: 0.4.0 cannot open the upgraded database. See upgrade and rollback guidance.
From ChatGPT to today's coding agents, interfaces have moved into IDEs, terminals, and browsers. We still tend to put our requests, questions, and feedback into a chat box.
As agents take on more work independently, the moments that need human involvement change: trying the result, adding context the agent lacks, or choosing between possible directions.
RambleDesk organizes interaction around those moments. It brings the current work, the questions that need your judgment, and your feedback together, giving each review a clear focus and a path back to the original task.
It suits a workflow where the agent works independently and you step in at meaningful points to try the result and give feedback.
---
config:
quadrantChart:
chartWidth: 800
chartHeight: 640
pointLabelFontSize: 16
xAxisLabelFontSize: 16
yAxisLabelFontSize: 16
themeVariables:
quadrant1Fill: transparent
quadrant2Fill: transparent
quadrant3Fill: transparent
quadrant4Fill: transparent
---
quadrantChart
title UI richness × extra token usage · AG-UI: connection layer
x-axis "Text chat" --> "Direct interaction"
y-axis "Fewer extra tokens" --> "More extra tokens"
"Classic chat": [0.10, 0.06]
"askUserQuestion": [0.29, 0.12]
"Chainlit": [0.18, 0.28]
"AI SDK": [0.33, 0.28]
"assistant-ui": [0.49, 0.28]
"CopilotKit": [0.65, 0.28]
"RambleDesk": [0.81, 0.28] radius: 7, color: #0ea5e9
"MCP Apps": [0.95, 0.28]
"Tambo": [0.40, 0.65]
"A2UI": [0.57, 0.65]
"json-render": [0.74, 0.65]
"OpenUI": [0.90, 0.65]
"OpenIntelligentUI": [0.71, 0.90]
RambleDesk applies the established prebuilt UI approach to work such as review, annotation, and hands-on trials.
Positions illustrate selected uses, not measured token costs; a project may span several positions. For an existing overview, see CopilotKit's Generative UI Spectrum. See related resources for project sources and more community maps.
Install Git, Node.js, pnpm, Rust, and your platform's Tauri dependencies, then run:
git clone https://github.com/l1veIn/rambledesk.git
cd rambledesk
pnpm install --frozen-lockfile
pnpm devSee development setup, agent setup and recovery, the workbench playground, or the documentation index.
- Review and annotation: recent refinements draw on Plannotator and Agentation.
- Related feedback workflows: MCP Feedback Enhanced is a close counterpart identified during later comparison.
- Brainstorming: Miro and FigJam demonstrate established card-grouping patterns; Luminate and Wordcraft explore human–AI creative work.
See sources and design connections for what informed these refinements and what is related work.
- Codeg: ACP integration, agent conversations, settings, and appearance references
- Snow Shot: the screenshot stack
- RepoChan: brand and character assets
- Kotone: the local speech stack this workbench grew from
MIT. Adapted code and other third-party components retain their respective licenses; see third-party notices.
