A free AI assistant,
and the knowledge to use it well
What is actually in the product
Two models, switchable per message
Use the fast model for quick questions and the larger model for reasoning and code, in the same conversation. Switching costs nothing and does not reset your thread.
Voice conversations
Speak your question and hear the answer read back. Useful for thinking out loud, for studying while your hands are busy, and for anyone who finds typing long prompts tiring.
Live artifact canvas
HTML, SVG and interface snippets render in a side-by-side preview as they are generated, so you can see the result and iterate without pasting into a separate editor.
Memory and custom instructions
Save your role, preferences and writing style once and they apply to later conversations. It is opt-in, and you can review or clear it in settings at any time.
Real-time web search
Ask it to search and it will read current pages before answering, which materially reduces invented facts. It can still misread a source, so the links are there for you to check.
Clear data handling
Conversations are stored so history persists and are never sold or used to train models. What that does and does not protect is spelled out plainly rather than implied.
An AI assistant, described accurately
PhantomAI is a browser-based AI assistant. You type or speak a question, or paste in the material you are working with, and it replies. Two models are available and you choose between them per message: GPT-OSS 120B for reasoning, code and longer answers, and GPT-OSS 20B when you want a fast reply to something short. Both are open-weight models served through Groq.
Around that sit the features that make a chat box into something you would actually keep open: a canvas that renders generated HTML and interfaces beside the conversation, real-time web search for questions that depend on current information, document and image analysis for material you upload, and memory so your instructions carry across conversations.
It is free. There is no paid tier and no card to enter — usage is metered with 150 daily credits, and every action's exact cost is published. PhantomAI is built and maintained by one person, Simanta Pratim Das, rather than a company.
The other half of this site is educational. Most people using AI assistants get mediocre results for reasons that are learnable, and most are never told where these tools reliably fail. So the guides are written to be useful whether or not you ever create an account here — including with a different AI tool entirely.
How an AI assistant actually produces an answer
This is worth two minutes, because almost every practical tip follows from it. A language model does not look up your answer. It breaks your text into tokens — word fragments — and repeatedly predicts which token most plausibly comes next, given everything in front of it. Then it does that again, with its own output now part of the input.
Three consequences fall straight out of that mechanism, and they explain most of what people find confusing:
Fluency and accuracy are separate
The model is optimised for plausible continuations, not verified ones. A false statement can be just as fluent as a true one, which is why a confident tone is not evidence and 'are you sure?' is not a check.
Giving it material beats relying on its memory
Facts recalled from training are reconstructed and can drift. Facts you paste into the conversation are right there in the input. This is the single biggest lever you have over accuracy.
Making it reason before concluding genuinely helps
Because each token is conditioned on the text so far, asking it to work through the problem before answering puts its own intermediate reasoning into the input it conditions on. Asking for the conclusion first gives it nothing to build from.
The same request, twice
Vague
Write about renewable energy.
Nothing here specifies audience, length, angle or scope, so the model averages over everything it has seen about renewable energy. The result is generic because the request was.
Specified
Write a 400-word section for a Year 10 geography revision guide explaining why solar power output varies by season in the UK. Audience: 15-year-olds preparing for an exam. Include: the angle of incidence, day length, and cloud cover. Avoid: cost, politics, and any statistics you are not confident about. End with three short recall questions and their answers.
Audience, length, format, three required concepts, explicit exclusions, and a defined ending.
Why the second one works
Every constraint eliminates a large set of plausible continuations. “Year 10 revision guide” fixes vocabulary and register; naming the three mechanisms stops it drifting into cost and policy; the exclusion of shaky statistics removes the most likely place for it to invent something; and the recall questions force it to commit to specific, checkable claims. You are not being polite to the model — you are narrowing what counts as a plausible answer.
Full prompting guideThree sequences that reliably work
Debug an error properly
- 01Paste the full error, not a paraphrase of it.
- 02Paste the function it points at, plus the call site.
- 03Ask: 'List three possible causes, ranked, and what I should check first for each.'
- 04Test the top cause yourself before asking for a fix.
Why it works: Asking for ranked hypotheses instead of a fix stops the model from committing to the first plausible story. You keep the diagnosis, which is the part it gets wrong most often.
Study something you half-understand
- 01Write your own explanation of the concept first, badly.
- 02Paste it and ask: 'What is wrong or missing in this explanation?'
- 03Ask for five questions that would expose a shallow understanding.
- 04Answer them without looking, then have your answers marked.
Why it works: Retrieval and correction build memory; reading a summary does not. Writing first also gives the model your actual misconception to work with instead of a generic topic.
Research without inventing sources
- 01Find the sources yourself — Scholar, PubMed, your library.
- 02Paste the abstracts or the relevant pages into the conversation.
- 03Ask it to tabulate the claims, methods and disagreements.
- 04Ask which conclusions the pasted material does not support.
Why it works: The model cannot fabricate a citation for a paper you supplied. Moving the sourcing step to a human removes the single most common failure in AI-assisted research.
Two models, and you pick
GPT-OSS 120B
Default — reasoning, code, long answers
openai/gpt-oss-120bGPT-OSS 20B
Fast — short questions and quick replies
openai/gpt-oss-20bMade for the way you work
Learn to use AI properly
What PhantomAI gets wrong
It can state things that are false, in confident and fluent prose. Tone tells you nothing about accuracy.
It will invent citations, papers and case law if you ask for references. Do not ask.
It is unreliable at arithmetic and counting — ask for a formula or a script instead of a number.
It cannot run or test the code it writes, so 'this works' is a prediction.
It has a knowledge cutoff, and search does not fully fix that.
It cannot act for you — no emails sent, nothing booked, no accounts accessed.
Where PhantomAI wins, and where it doesn't
Designed to stay out of the way
Responses stream as they generate, models switch without leaving the thread, and saved prompts turn repeated multi-step tasks into one click. The goal is fewer small frictions, not more features.
Security, stated plainly
Authentication is handled by Clerk and all traffic is served over HTTPS. There is no dedicated security team and no SOC 2 or ISO 27001 certification — this is a solo project, and claiming enterprise-grade assurances would be dishonest.
What happens to your data
Conversations are stored so history works, and are sent to Groq to generate replies. They are not sold and not used to train models. They are not end-to-end encrypted. You can delete conversations, or request account deletion by email.
Built on a modern, boring stack
Built in the open, by one accountable person
Open about how it's built
PhantomAI is a solo-built product — no anonymous team, no invented case studies. The founder's real GitHub profile is linked from the About page.
A real person behind it
Simanta Pratim Das builds and maintains PhantomAI end to end from Guwahati, India, and reads every message sent through the contact page personally.
Published limits, not just features
There is a dedicated page listing what PhantomAI cannot do, and the guides spend as much time on where AI fails as on what it is good for.
A few quick answers
What is PhantomAI?
PhantomAI is a free AI assistant you use in the browser. It supports two switchable models, voice conversations, a live canvas for rendering code, real-time web search, and memory for your custom instructions. The site also publishes free guides on using AI assistants well.
Is PhantomAI free to use?
Yes, and there is no paid plan. Every account gets 150 credits a day — a fast-model message costs 5, the default model costs 10, search adds 4. The exact cost of every action is published on the pricing page.
Which AI models does it use?
Two open-weight models served through Groq: GPT-OSS 120B as the default for reasoning and code, and GPT-OSS 20B for fast short answers. PhantomAI does not run GPT-4, Claude or Gemini, and does not train its own models.
Can the answers be wrong?
Yes. It predicts plausible text rather than retrieving verified facts, so it can be confidently wrong — especially with citations, arithmetic and recent events. There is a dedicated limitations page, and a guide on how to fact-check AI output quickly.
Who builds PhantomAI?
Simanta Pratim Das, an independent developer in Guwahati, India. It is a solo project — no company, no support team. The About page links a verifiable GitHub profile.
Is my data private?
Conversations are stored so your history persists, and are sent to Groq to generate responses. They are not sold and not used to train models. They are not end-to-end encrypted, because a model has to read your text to answer it. Do not paste passwords, keys or other people's personal data.
Start with a guide,
or start with a conversation
The guides are free and need no account. The assistant is free too — 150 credits a day, no card, nothing to cancel.