RightNow AI
RightNow AI is an AI-powered code editor built specifically for NVIDIA GPU kernel development, combining CUDA, Triton and Mojo editing with real-time Nsight profiling, automated benchmarking, a GPU emulator covering dozens of architectures, and remote execution over SSH.
What is RightNow AI?
RightNow AI, Inc. presents itself as a GPU research lab rather than a single product. Three things sit under the name: the RightNow Editor, Forge, and RunInfra, which lives on its own domain. The editor is the part most people meet first.
It is a desktop application, rebuilt on VS Code with the 1.0.0 release of 4 February 2026, and it exists for one job: writing, profiling and tuning GPU kernels. It runs on Windows, macOS and Linux, on x64 and ARM alike, and expects a CUDA-capable NVIDIA GPU with the CUDA Toolkit 11.0 or later. You can write in CUDA, Triton, CUTE, TileLang, Numba, Mojo, CUDA Tile or PyTorch.
Profiling is where the tool makes its case. It drives the NVIDIA Nsight Compute CLI under the hood, surfaces SM efficiency, memory throughput, occupancy and cache hit rates as inline CodeLens metrics, and opens the full Nsight interface in one click with your profile already loaded. A PTX and SASS viewer shows what the hardware really executes. Benchmarking sweeps block sizes, tile sizes and memory layouts automatically. A GPU emulator covers 86 or more architectures so you can test on an A100 or H100 you do not own, and remote cards are reachable over SSH.
On the AI side, four cloud models are served through a proxy, priced in credits: Claude Sonnet 4.5, Claude Haiku 4.5, GPT-5 Codex and DeepSeek Chat V3, with separate fill-in-the-middle models for autocomplete. Ollama, vLLM and LM Studio cover the offline path, where the code never leaves the machine. Custom agents with skills and MCP integrations arrived with the same 1.0.0 release.
Forge is the enterprise arm: a swarm of 32 parallel coder-and-judge agent pairs that turns a model into optimised CUDA or Triton kernels, delivered as a drop-in replacement and verified for numerical correctness. It targets datacenter cards such as the B200, H200, H100, L40S and A100, and is sold on custom pricing only. A Forge command-line tool installs through npm. The company is backed by Y Combinator and publishes both arXiv papers and open-source repositories.
What it does
- Profile CUDA kernels in real time inside the editor, driving NVIDIA Nsight Compute underneath
- Pinpoint the bottleneck and get a fix suggested from the measured profiling metrics
- Write and complete GPU code with a model that knows the target architecture
- Emulate more than 86 GPU architectures without owning the hardware
- Benchmark block sizes, tile sizes and memory layouts automatically, then compare the results
- Inspect the PTX and SASS the GPU actually executes, line by line
- Run kernels on remote GPUs over SSH, on cloud instances or on-premise servers
When to use RightNow AI / When not to
A quick filter to help you decide if RightNow AI is the right fit.
When to use RightNow AI
- CUDA kernel engineers who profile, tune and ship low-level GPU code every day
- ML and inference teams trying to cut GPU spend by optimising the kernels behind their models
- Developers without access to A100 or H100 hardware, who need an emulator to test before they deploy
- Academic researchers and PhD students stuck waiting in GPU cluster queues
- Enterprises that require on-premise deployment, a custom SLA, an NDA and IP protection
When not to use RightNow AI
- Anyone working on AMD, Intel or Apple GPUs, since the editor targets NVIDIA CUDA hardware only
- Developers without a CUDA-capable NVIDIA GPU and the CUDA Toolkit 11.0 or later installed
- Teams that want a mobile app, as the tool ships only as a Windows, macOS or Linux download
- Engineers who need to drive the product through a public API, because none is documented
- Organisations bound by HIPAA, FISMA or the GLBA, which the terms of use explicitly exclude
How to use RightNow AI
A typical end-to-end flow, from setup to results.
- Install the NVIDIA CUDA Toolkit 11.0 or later, which the editor requires before it can profile anything
- Download the build matching your operating system and architecture from the downloads page
- Create an account on the dashboard; the free plan needs no payment details
- Note that cloud models are pre-configured through RightNow Proxy, so there is no API key to paste
- For offline work, point Settings then AI Providers at a local Ollama, vLLM or LM Studio endpoint and test the connection
- Choose your autocomplete model under Settings, AI Providers, Autocomplete Model
- Open or write a kernel in CUDA, Triton, CUTE, TileLang, Numba, Mojo or PyTorch
- Profile it from the profiling terminal, read the metrics in CodeLens, or open the full Nsight interface in one click
- Benchmark configurations, compare GPUs side by side, or paste SSH details to run on a remote card
- For Forge, install the CLI through npm, run forge login, then forge to launch the interactive wizard
Pros & Cons
Pros
- A permanent free plan with unlimited profiling and benchmarking, not a countdown trial
- The GPU emulator removes the need to buy or rent an A100 or H100 simply to test a kernel
- Local LLM mode keeps source code on the machine, with nothing sent to a third party
- The terms state plainly that customer code is never used to train models and stays the customer's property
- Ten subprocessors are named with their role and location, which is rare transparency at this size
- Retention and deletion rules are documented with actual figures rather than vague language
- An entry price of 20 USD per month is modest against the cost of the hardware it targets
Cons
- NVIDIA hardware only, with no AMD or Intel support announced anywhere on the site
- The site contradicts itself on API keys: the pricing FAQ says you can bring your own, the documentation says there is no BYOK
- Emulator claims vary by page, from 86 or more architectures and under 2% error to 50 or more GPUs and 98% accuracy
- Forge carries no public price, is enterprise-only, and is reachable solely through a sales form
- All purchases are non-refundable and subscriptions renew automatically each month
- Hosting is United States only, with no EU option, no Article 27 representative and no DPA on offer
- The SOC 2 Type 2 badge links to a trust center for RunInfra, a sibling product, showing no published policies, subprocessors or documents
Pricing & Plans
A permanent free plan is available at no cost, with unlimited profiling and benchmarking. The cheapest paid tier is Pro at 20.00 USD per month, billed monthly in US dollars, with Visa and Mastercard accepted. Forge, the enterprise kernel optimisation service, is offered on custom pricing through a sales contact. All purchases are non-refundable.
- single-GPU development
- unlimited profiling and benchmarking
- CodeLens performance metrics
- GPU virtualisation
- automatic kernel fusion
- local LLM support
- limited autocomplete
- one Forge credit per month and community support on Discord. The GPU emulator is not included.
- everything in Free
- plus GPU emulator access to more than 50 GPUs
- multi-GPU comparison up to six cards
- natural language profiling
- unlimited autocomplete
- GPU-optimised suggestions
- 1000 AI Agents credits and one Forge credit per month
- and priority email support. Multi-GPU execution analysis is marked as coming soon.
- AI kernel optimisation across all GPU architectures
- drop-in replacement kernels
- dedicated infrastructure
- on-premise deployment
- a custom SLA
- NDA and IP protection
- a dedicated support team and unlimited multi-GPU work.
- A Mega tier is referenced in the AI providers documentation but does not appear on the pricing page.
Data, GDPR & hosting
A consolidated view of how RightNow AI handles your data.
GDPR overview
Concrete, and better documented than the size of the company would suggest, but incomplete. The privacy policy, last updated on 24 December 2025, sets out four lawful bases for processing: contract performance, legitimate interests, consent and legal obligation. A dedicated data subject rights section covers access, rectification, erasure, portability as a JSON export, restriction, objection and withdrawal of consent, with a 30-day response commitment and jaber@rightnowai.co as the point of contact. International transfers rely on Standard Contractual Clauses with every vendor, plus encryption in transit and at rest. Three gaps stand out: no Article 27 EU representative is named, no data protection officer is designated, and no Data Processing Agreement is mentioned anywhere on the site.
Who owns the data?
Section 21 of the terms of use is unusually clear. You keep all ownership rights, title and interest in your source code, CUDA kernels, GPU profiling data and anything else you upload or create. RightNow AI, Inc. claims no ownership and processes that material solely to deliver the optimisation and profiling services. It shares the material with third parties only where the service requires it, OpenRouter being the example the company itself gives, and treats it as confidential, disclosing it only to run the service, with your explicit permission, or where the law compels it. Feedback and suggestions are the exception: their intellectual property is assigned to the company.
Reuse rights
You can reuse your own material freely, without asking. The terms confirm that AI-generated optimisations and suggestions are handed over to you and that you retain all rights to the optimised code, which Forge delivers as a drop-in replacement exposing the same API. The restriction runs the other way round: the site's own content, trademarks and logos are licensed to you only for personal, non-commercial use or an internal business purpose, and systematically retrieving data from the service to build a collection, database or directory requires written permission from the company.
Data retention & training
Hosting summary
All hosting is in the United States, with no European option. The privacy policy states that the services are operated from the US and that all data is stored in US-based data centres: the primary database on AWS us-west-1 in Oregon, the API infrastructure in Chicago, Illinois, and the vector database on GCP us-central1 in Iowa. Ten subprocessors are named, all US-located: Supabase for database and authentication, OpenRouter for AI processing, Turbopuffer for vector search, Upstash for rate limiting and sessions, Stripe for payments, Resend for email, Amplitude for analytics, Sentry for error monitoring, Fly.io for API hosting and Vercel for web hosting. For transfers, the company relies on Standard Contractual Clauses with every vendor, encryption in transit and at rest, and infrastructure it describes as SOC 2 certified, each provider holding its own certifications. EU users are told they consent to the transfer by using the service and may complain to their local supervisory authority.
Things to keep in mind
Risks and trade-offs to weigh before adopting RightNow AI.
- In cloud mode your source code and kernels are sent to a third party, OpenRouter, before any optimisation comes back, so check how sensitive that code is first
- The terms disclaim all warranties on AI-generated optimisations: you remain responsible for reviewing, testing and validating every suggestion before production
- Emulator results are approximations by nature, and the site quotes different accuracy figures on different pages
- Leaning on the AI to read profiler metrics can quietly erode the low-level skill that makes those metrics meaningful in the first place
- All data lives in the United States, so any European user consents to that transfer simply by using the service
- Purchases are non-refundable and subscriptions renew automatically every month unless cancelled
- Ideas, suggestions and feedback sent to the company count as Submissions, with their intellectual property assigned away
Setup & Integrations
Technical difficulty
Moderate, and front-loaded on hardware rather than software. You need a CUDA-capable NVIDIA GPU and the CUDA Toolkit 11.0 or later installed separately; the editor itself is a straightforward download per operating system and architecture. Cloud models are pre-configured through a proxy, so there is no API key to obtain. The more technical paths are optional: pointing the editor at a local Ollama, vLLM or LM Studio server, or installing the Forge CLI through npm. The intended audience already knows CUDA and Nsight Compute, and the product assumes it.
Deployment
Integrations
Supported languages
Behind RightNow AI
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement RightNow AI.
Frequently asked questions
Which GPUs does RightNow AI support?
Is there a free plan, and what does the first paid tier cost?
Is my source code used to train AI models?
Can I keep my code off the network entirely?
Where is my data hosted?
How long is my data kept, and how do I delete it?
What do I need before installing the editor?
Can I bring my own API keys?
Is there a public API or a mobile app?
What is the minimum age to use the service?
Should you pick RightNow AI?
RightNow AI is a narrow tool that knows exactly how narrow it is. It will not help you build a web app; it exists so that engineers writing CUDA kernels can profile, benchmark, emulate and tune them without leaving the editor, and on that ground it is unusually complete. NVIDIA Nsight Compute runs underneath rather than being replaced, which is a sensible position, and the emulator answers a real problem: testing on an H100 you do not own.
The legal posture is better than the age of the company would suggest. The terms state that you keep ownership of your code, that it is never used to train models, and that it is deleted within thirty days of account closure. Ten subprocessors are named with their role and location, and retention is quantified rather than merely described. Very few tools this young publish that much.
The weaknesses are mostly about consistency and reach. Hosting is entirely American, with no EU residency option, no Article 27 representative and no Data Processing Agreement, which is a real obstacle for European organisations. The site contradicts itself on whether you can bring your own API keys, quotes three different sets of emulator figures, and references a Mega plan the pricing page does not sell. The SOC 2 Type 2 badge links to a trust center belonging to a sibling product rather than to this one, which is the kind of detail a procurement team notices.
The barrier to entry is zero on price and high on hardware: the free plan is permanent and genuinely usable, but you still need an NVIDIA GPU and the CUDA Toolkit. For anyone already in that position, evaluating it costs nothing.
- Choosing a selection results in a full page refresh.
- Opens in a new window.