Introducing the Decuff Web API: a browser on Tor for your AI agent.
→ Read breach reports at the source
→ Post to whistleblower sites through Tor
→ Search 9,000 onion sites
→ Hide where your traffic comes from
Zero retention by default. No KYC.
decuff.ai/reference
One prompt. A working app. Real run.
decuffed-flagship wrote a single-file expense tracker: form, running total, canvas donut chart, saved to localStorage.
9.24s · 5,478 tokens · 760 lines · $0.0073
We clicked through it in Chrome: 0 console errors, 0 edits. It survives a reload.
Describe it. Run it. decuff.ai
I asked our flagship model for a full app in one prompt and didn't touch the code.
An expense tracker with a form, delete buttons, a running total and a donut chart drawn without any libraries, saved in the browser.
760 lines in 9.24 seconds, for less than one cent ($0.0073).
We opened it in Chrome, added expenses, deleted one and reloaded. Zero console errors and zero edits.
decuff.ai
One prompt. A working app. Real run.
decuffed-flagship wrote a single-file expense tracker: form, running total, canvas donut chart, saved to localStorage.
9.24s · 5,478 tokens · 760 lines · $0.0073
We clicked through it in Chrome: 0 console errors, 0 edits. It survives a reload.
Describe it. Run it. decuff.ai
Streaming vs not. Same prompt, 5 runs each, decuffed-flash.
→ stream: true: first words in 1.12s (median)
→ stream: false: nothing until 5.60s
Total time is about the same (5.42s vs 5.60s). Your user just starts reading 5× sooner.
It's one flag: stream=True. decuff.ai/docs
1M free tokens. How far do they actually go?
We measured real tasks on decuffed-flash:
→ quick question: 145 tokens (≈6,896 per 1M)
→ write a function: 304 (≈3,289)
→ review a code file: 1,397 (≈715)
→ 4-turn chat: 7,029 (≈142)
Chats cost more each turn because the whole history is sent again.
Start free with a phone sign-in: decuff.ai/models
One Decuff key. Three clients. Real runs.
Same prompt, same endpoint: api.decuff.ai/v1
→ OpenAI Python SDK: 2.95s · 23 in / 43 out
→ curl: 2.37s · 23 in / 46 out
→ Node fetch, no SDK: 2.37s · 23 in / 44 out
3/3 worked. Total: $0.0003.
Change the base URL. Keep building: decuff.ai/docs
We hid one sentence in a web page. Our support bot emailed its secret to a stranger.
Controlled prompt-injection test on our own demo app (fictional shop, fake staff code), running decuffed-flash:
→ asked directly: leaked 3/10
→ hidden in a page it was asked to summarize: leaked 10/10
→ after the fix: 0/10
The fix: fence untrusted text, allowlist tool recipients, scan output for secrets.
Same page on decuffed-flagship: 0/10.
Attack yourself first. decuff.ai/breach
One dark prompt. One real run.
Gothic horror: a drowned keeper knocks on his widow's door.
decuffed-flagship wrote it start to finish:
"Maren. Let me in. I'm so cold."
0 refusals · 0 disclaimers
15.09s · 243 tokens · $0.00035
Fiction and dark themes are allowed by design. Every wall is listed → decuff.ai/acceptable-use
Free trial on every account → decuff.ai
If your app has an LLM in it, test it like an attacker would.
OWASP ranks prompt injection #1 on its Top 10 for LLM apps. Their own guidance says it's unclear any method fully prevents it.
Text from a user, a webpage or a file can steer the model. Wire it to tools and data, and that steer becomes an action.
Filters help. Adversarial testing shows what actually gets through.
Our Breach does that: authorized AI pentesting on systems you own, only in the scope you set.
Early access → decuff.ai/breach
One dark prompt. One real run.
Gothic horror: a drowned keeper knocks on his widow's door.
decuffed-flagship wrote it start to finish:
"Maren. Let me in. I'm so cold."
0 refusals · 0 disclaimers
15.09s · 243 tokens · $0.00035
Fiction and dark themes are allowed by design. Every wall is listed → decuff.ai/acceptable-use
Free trial on every account → decuff.ai
Same prompt. Two models. Real run.
Task: build a working analog clock in a single HTML file.
⚡ decuffed-flagship: 2.73s · 1,167 tokens · $0.0016
🧱 decuffed-flash: 49.11s · 3,934 tokens · more detailed design
Both worked on the first try.
Pick speed or detail. Try both free → decuff.ai
Not every model is right for every job, including ours.
On this task, Flagship finished in under 3 seconds. Flash took 49.
Flash is a smaller 27B model that writes more code and more detail, so it trades speed for depth.
Our guidance:
→ Flagship: speed and 1M context
→ Flash: richer output when time isn't the constraint
We'll keep publishing real benchmarks, including the ones where we're slower.
Same prompt. Two models. Real run.
Task: build a working analog clock in a single HTML file.
⚡ decuffed-flagship: 2.73s · 1,167 tokens · $0.0016
🧱 decuffed-flash: 49.11s · 3,934 tokens · more detailed design
Both worked on the first try.
Pick speed or detail. Try both free → decuff.ai
build your own ai chatbot in just 31 lines of python using decuff.
→ powered by decuffed-flash
→ simple openai-compatible api
→ supports streaming responses
→ maintains conversation history
→ 1M free tokens every month
all you need is python and the openai package.
no complicated setup. just build and start chatting.
try it out: decuff.ai
this is the setup i'd recommend to any developer trying decuff.
no new tools or complicated workflow. just claude code, an api key and a base url.
decuffed-flash built this working app in 37 seconds from a single prompt.
give it a try on your own project and let me know how it performs.
build your own ai chatbot in just 31 lines of python using decuff.
→ powered by decuffed-flash
→ simple openai-compatible api
→ supports streaming responses
→ maintains conversation history
→ 1M free tokens every month
all you need is python and the openai package.
no complicated setup. just build and start chatting.
try it out: decuff.ai