About · Graytell Labs
Trustworthy AI,
built in the open
Founder
Anubhav Sapkota
Graytell Labs
Based in
Fujairah
Fujairah, United Arab Emirates
Licenses
Apache-2.0 · MIT
Everything, always
What we build
Three products, one standard
Each one exists because we needed it for the next — the registry renders the agent, the agent produces the research, the research sharpens the standard.
Graytell Agent
The open research agent. Multi-provider routing, live tool use, and full citations — ask a question, watch every step, check every source.
Apache-2.0 · TypeScript
Gray UI
The canvas for your next interface — 168 copy-paste React components in a strict monochrome design system. 55 shadcn/ui primitives, all 48 AI elements, 12 agent blocks.
MIT · shadcn-style registry
Impersio
Where Graytell began: a dictation-first deep-research agent with 30+ models across 5 providers, real-time tool calling, and streaming answers.
Live · the lab's first product
The story so far
From a dictation app to a research lab
Graytell started as Impersio — a dictation-first AI agent with real-time tool calling and 30+ models behind one interface. Shipping it raised a question that would not go away: users could not tell which answers to trust.
So the lab went looking for the foundations. That work became the Graytell research agent — a research-first agent where every tool call is visible and every claim carries a citation — and the Citation Integrity benchmark, which measures how faithfully frontier models actually cite their sources under agent conditions.
Building the agent's interface produced a third artifact: Gray UI, the strict monochrome component registry this site runs on — because an AI safety lab should not need a rainbow to be readable.
Everything we publish — the agent, the registry, the paper, the scored data — is open source under Apache-2.0 or MIT. If it matters enough to claim, it matters enough to inspect.
How we work
Rules we do not bend
- Safety is a spec
- Trustworthy AI is not a promise, it is a workflow: citations by default, permissions fail-closed, and benchmarks we publish ourselves — including when the numbers are unflattering.
- Open by default
- The agent, the registry, the paper, the data. Apache-2.0 and MIT, because foundations nobody can inspect are not foundations — they are claims.
- Small lab, real artifacts
- We ship things you can use today: a research agent you can run, a component registry you can build on, and research you can cite.
Say hello
The fastest way to reach the lab is our Discord — the founder answers there personally.