Sanket Dongre / TinkerClaw
I'm building the next interfaces for working with agents.
TinkerClaw is my public workbench for turning interface questions into shipped products, open-source tools, and field notes. I'm exploring what comes after the chat box: voice, ambient reading, screen-aware controls, generated UI, XR, and mixed-reality workspaces.
Builder note
From investigations to products.
I use this site to document the questions, prototypes, failures, and releases behind my agent-interface work. The output should be inspectable: essays, working demos, open repos, and products people can actually use.
What will future agent interfaces feel like, and how do we know when they work?
Generative UI for Agents
Interfaces that emerge from model output: forms, panels, timelines, overlays, and task surfaces for agents when chat is the wrong shape.
Voice and Ambient Interfaces
Voice, narration, listening, interruption, and background presence as first-class ways to interact with agents. Scroll is the first example.
Spatial, XR, and Mixed Reality
Experiments where agents create or control interfaces across screens, rooms, headsets, simulations, and mixed-reality workspaces.
Evaluation and Trust
Rubrics, traces, and field notes for deciding whether new agent interfaces are clearer, safer, and less cognitively expensive.
P-001
Scroll
Research question
Can a voice-based, screen-aware agent adapt to reading velocity in real time, and reduce cognitive load versus manual LLM querying?
Current artifact
Scroll for Mac: ambient voice narration, velocity-aware summarization, real-time Q&A over screen content, and a one-time purchase model.
What it tests
Screen OCR pipelines / velocity heuristics / low-latency voice / ambient UI / attention alignment
Reading context: active
Narrating
Velocity: 2.4x
Exploration areas.
Generative UI for Agents
How should agents decide when to answer with text, when to generate an interface, and what shape that interface should take?
Voice-Based Agent Interfaces
How should people talk to, interrupt, guide, and listen to agents? Scroll explores voice as an ambient reading interface, not just a command channel.
Ambient and Screen-Aware Interfaces
What happens when agents understand what is visible, where your attention is, and what you are doing before you explicitly ask?
XR and Mixed-Reality Agent Interfaces
How should agents appear in headsets, rooms, simulations, and spatial workspaces without adding clutter or cognitive load?
Human-in-the-Loop Agent Interfaces
How should humans inspect, interrupt, approve, roll back, and collaborate with AI systems that can act on their behalf?
Evaluation, Safety, and Observability
How do we measure agent interface quality, cognitive load, traceability, uncertainty, and failure recovery across modalities?
Artifacts and products.
The current tools stay intact. They now sit beside the research as shipped outputs, experiments, and commercial artifacts from broader agent-interface exploration.
Scroll
The first shipped artifact from Investigation #1. Ambient, voice-based, screen-aware AI narration for macOS. It reads what is on screen, adapts to how fast you move, and answers when you ask.
$49 one-time
Visit getscrollapp.com
Claw Agents
34 pre-built agents that extend OpenClaw with email, content, ops, finance, and research workflows. Install the pack you need with a single command.
$49 lifetime
Browse agents →
Wrapper Kit
A complete starter kit for launching your own OpenClaw wrapper product: landing page, billing, onboarding, and deployment scripts ready to white-label.
$49 one-time
Learn more →
Field notes and research log.
A public notebook for observations, essays, and investigation updates as the work happens.
Scroll as Investigation #1
A working investigation into ambient and voice-based agent interaction: screen context, scroll velocity, narration, attention alignment, and real-time questions.
Read →The Case for Human-Factors-Grade Agent Interfaces
Why future agent interfaces need more than pretty components: legibility, cognitive load, trust calibration, and reproducible evaluation.
Read →When Should an Agent Generate an Interface?
A decision tree for moving from chat answers to generated forms, timelines, maps, workspaces, and spatial control panels.
Read →Open artifacts.
Repos, rubrics, examples, and components released from the work. Enough is open to make the thinking inspectable.
vibe-building-skills
Open-source agent skill library for product strategy, marketing systems, creative generation, frontend design, and reusable building workflows.
open-model-council
Perplexity-style model council for the web: multiple LLMs draft, debate, judge, synthesize, and preserve per-model reasoning trails.
loop-codex-plugin
Codex plugin for loop engineering: project loops that discover, triage, act, verify, report, and remember with durable state.
codexbridge
Unity Editor bridge for Codex and Claude Code: in-editor chat, Unity-aware tools, project context presets, saved chats, export, and optional Meta XR tool exposure.
OpenXRSim
Unreal Engine OpenXR simulator for testing XR gameplay on desktop with simulated HMD, controllers, rooms, and record/replay timelines.
About me.
I'm Sanket Dongre. I build TinkerClaw as a public workbench for future agent interfaces: voice, ambient, screen-aware, XR, mixed reality, and Generative UI.
My background sits at the intersection of human-factors visualization, spatial computing, AI agent architecture, and product prototyping. The thread through the work is simple: how should people understand, guide, interrupt, and trust increasingly capable agents?
Scroll is the first concrete investigation in that direction. It uses voice, screen context, and reading behavior to explore a more ambient way to work with AI without forcing everything through a chat box.