Thesys retweeted
Jun 13
Someone just open-sourced a new standard for AI-generated UI.
OpenUI lets your LLM stream real components (charts, forms, dashboards) directly into React. No more parsing massive JSON blobs.
→ 67% fewer tokens than JSON
→ 4.9s render vs 14.2s for Vercel's approach
→ Web, mobile, and email support
100% open source.
2
4
76
Andha kanoon retweeted
Got a chance to represent @thesysdev and OpenUI at the Agent Harness Hackathon as a sponsor judge! Amazing to see brilliant use cases - from FDA drug analysis to fun projects like using OpenUI to make music!
AWS loft served some amazing 🍕
1
8
191
Jun 13
what is under the hood? copilotkit or arrowjs or openui or thesysdev?
2
2
17
Jun 12
@metrics_co - the folks powering portfolio analysis for 150 VC/PE firms, GC, Accel, 8VC - just wrote a blog post on how they shipped Generative UI within their AI Analyst and why they chose OpenUI over JSON / Raw HTML
The blog post is a treasure trove of information with good architectural patterns on how they integrated Generative UI capabilities within their agent that already shipped markdown without breaking backward compatibility
Go read!
standardmetrics.io/from-text…
3
9
97
Jun 11
your llm and you when ur NOT using OpenUI
Jun 11
Most debates about Generative UI are actually two decisions being confused as one.
1. Transport: where does the UI appear? (AG-UI for your app, MCP-Apps for ChatGPT/Claude)
2. Generation: what does the agent emit? (Static, Declarative, or Open-Ended)
Any pairing is valid. Once you separate them, most disagreements dissolve.
Wrote a full report covering the protocols, the formats, the token benchmarks, and everything you should consider before choosing your stack.
4
51
Jun 11
What it covers:
1/ AG-UI vs MCP-Apps: not a choice, two different answers to two different questions
2/ Static vs Declarative vs Open-Ended generation: where each one wins
3/ Token benchmarks: OpenUI Lang vs A2UI vs json-render vs raw HTML
4/ A decision framework for choosing your stack
Biggest takeaway:
- Transport is not really a choice. Once you chose an agent surface (Claude plugin vs in-product) it becomes a single option
- Format choice is a latency and cost choice. At 60 tok/s, the same dashboard renders in 19s as OpenUI Lang vs 41s as JSON patches.
openui.com/blog/state-of-gen…
5
147
Jun 11
Really proud of these contributions.
Huge thanks to the team and maintainers for building an amazing project and fostering such a great community.
If you haven't checked out @thesysdev 's OpenUI yet, you're missing out.
1
4
59
Thesys retweeted
Jun 9
It's really great to see my project AI Native Dashboard get listed in OpenUI Projects.
Thanks to @thesysdev team. It is really cool to explore UI generation using natural language
ALT https://www.openui.com/projects
1
1
6
305
You collected gigabytes of useless trash on your PC/VPS/VDS
Over roughly the last six months, the AI industry made a lot of big jumps. Tools changed, models changed, approaches changed. The last one changed 180 degrees.
Old personal orchestrations and first skills across three agents turned into a cleaner structure. The term harness, ready tools, agents with personality, all that stuff.
And then I opened my old files.
Dozens of projects. A hundred GB. uv/pip and Docker containers leaving cache everywhere. I had 40 GB of trash there.
Then local models: 4, 8 and 28 GB each. WebUI, OpenUI, Comfy. Pinokio cache after local TTS installs. Another ~40 GB of outdated files that had not been launched for 3 months.
But GBs are not the worst part. The flow was outdated too.
Dead skills, scripts, memory files and old paths inside instructions. Old orchestration plugin I dont use anymore because GSD, gstack and different paperclip things exist now.
The biggest pain is stale context.
Duplicate folders inside the LLM-wiki because an agent went insane and created video/research notes in random places. Three different skills for one task because Claude misunderstood me and created a new skill instead of updating the old one.
I cleaned it step by step, rewrote the useful parts into the LLM wiki and deleted the rest to zero. Ran agents in loops on tasks and self-diagnostics just to see the final flow.
This is exactly why Hermes is nice. It understands its own skill mess, updates things on the fly and creates fresh stuff only when it actually needs to.
Here is the prompt to fix this:
ROLE
Disk audit diagnostic agent. INSPECT and WRITE A REPORT, do not clean.
IRON RULE
Do not delete/move/rename anything. READ-ONLY only: du, df, ls, find, stat, docker system df.
NO cleanup commands: rm, docker prune, uv/pip cache clean, git gc.
If doubtful – flag it in the report, do not act.
WHAT TO LOOK FOR
1. Caches: uv, pip, npm, Docker orphans, build cache, dangling volumes, stopped containers, Pinokio.
2. Models/ML: weights .safetensors/.ckpt/.gguf/.bin, ComfyUI/WebUI/OpenWebUI models/output/venv, old untouched files.
3. Dead projects: unchanged 90 days; node_modules, .venv, build/dist, __pycache__, dumps.
4. Duplicates: folders/files by name and content, copies of models/datasets.
5. Outdated context MOST IMPORTANT: unused or duplicate skills, dead memory files, outdated instruction paths, obsolete plugins.
REPORT
Table with Path | Size | Last used | Confidence HIGH/MED/LOW | Why | Recommendation.
Recommendation means command for the user, not an action.
Verify usage, dont guess. If you cant verify, write “could not verify”.
End with these sections:
Total GB that can be freed
Top 5 quick wins
DANGEROUS – ask user first
Decision to delete belongs to the user.
1
3
162
Found this repo, it seems like it generates UI with 67% fewer tokens than JSON. Thoughts on this?
-/github.com/thesysdev/openui
2
25
Jun 4
Very cool! I also like openui, especially the reduced token syntax. I'm sticking with json-render for a bit more as it has really good action/loader handlers.
Also. We gotta meet. I'm finally staying put with no travels. Will ping you in the coming week or two
1
3
58