Forge CLI
Forge CLI generates drop-in CUDA C++ or Triton kernels tuned to your datacenter NVIDIA GPUs from a PyTorch or HuggingFace model ID.
If your inference bill is big enough that a 3-5x kernel-level win changes the cost model, Forge deserves a trial on your own checkpoint — the named per-model speedups and the 97.6% correctness pipeline are specific enough to test rather than trust. Pin down the per-kernel credit cost before you commit, since Forge is billed in credits while RightNow's published pricing page covers coding plans and Model APIs. Also note where the roadmap sits: the changelog since September 2026 is all RightNow
Verified 2d ago · liveness 80/100 · cite: rightaichoice.com/tools/forge-cli
- ML platform teams optimizing LLM or diffusion inference on H100, H200, B200, A100 or L40S capacity
- Infrastructure engineers measured on GPU utilization and cost per token
- Teams that need numerical correctness verification, not just a faster kernel
- Organizations with in-house PyTorch or HuggingFace models they cannot rewrite by hand
- Developers on consumer or RTX GPUs — this targets datacenter NVIDIA hardware
- Teams wanting a GUI or no-code optimization workflow
- Anyone needing sub-minute, interactive kernel tuning rather than a batch run
We scan live Reddit threads, YouTube comments, X posts, G2 reviews and other communities — and hand you an honest verdict in under a minute.
- Honest verdict, not marketing
- Real pros & cons from real users
- Attributed quotes with receipts
3 free scans · no card needed
Skip Forge CLI if you're tuning on consumer or RTX GPUs, need an interactive sub-minute loop rather than a batch run, or already clear your throughput target with torch.compile and a tuned vLLM setup.
Forge runs on a credit system — 1 credit per KernelBench or custom kernel, 1-2 credits for HuggingFace models — so a wide benchmark sweep across many checkpoints burns credits faster than a single-model run.
Forge sits in a different buying motion from RightNow's published self-serve tiers. Those are RunInfra coding plans — $10/mo Starter, $29/mo Pro, $99/mo Team — plus pay-as-you-go Model APIs from a $5 top-up, with Enterprise on custom terms including SOC 2 Type II and SLAs up to 99.99%. Forge kernel generation runs on its own credit system (1 credit per KernelBench or custom kernel, 1-2 for HuggingFace models), and no Forge credit rate is published on the pages we reached. The economics only
In short
Forge CLI — Forge CLI generates drop-in CUDA C++ or Triton kernels tuned to your datacenter NVIDIA GPUs from a PyTorch or HuggingFace model ID. Best for ML platform teams optimizing LLM or diffusion inference on H100, H200, B200, A100 or L40S capacity, Infrastructure engineers measured on GPU utilization and cost per token, Teams that need numerical correctness verification, not just a faster kernel. Plans from $5.
What's new in Forge CLI
Checked todayAcross the latest 3 updates: 1 launch and 2 changelog entries.
RightNow Agent 1.0.4 ships native installers and in-app updates across six platforms
First hosted release adds binaries for Windows x64/ARM64, macOS Apple Silicon/Intel and Linux x64/ARM64, with SHA-256 sidecars, GPG-signed Linux builds and a manifest-verifying updater.
RightNow Agent 1.0.5 fixes Windows console paste handling and model list refresh
Windows ConPTY multi-line pastes fold into one chip; /model list refreshes on late catalog arrival; Windows ARM64 artifacts signed with refreshed checksum.
RightNow Agent 1.0.3 ships legacy Windows x64 zip distribution
Zip release installs under %LOCALAPPDATA%\Programs\RightNow; superseded days later by the 1.0.4 native installer.
What people actually say about Forge CLI — is it worth it?
We ran a structured research pass across product reviews, community discussions, and post-purchase forum threads to surface the patterns vendors won't publish themselves. Below: the recurring strengths, the hidden costs people mention most, and the cohort that consistently regrets adopting this tool.
34 mentions across 5 sources (Hacker News, YouTube, Product Hunt, GitHub, Lemmy) · researched Aug 14, 2026.
Average across the 5 sources that answered — each source counts once, not each post.
- +Delivers 3-10× speedups over torch.compile for LLM inference.
- +Automates CUDA/Triton kernel generation, saving manual tuning effort.
- +100% numerical correctness verification via tiered evaluation.
- +Supports all NVIDIA datacenter GPUs, including B200 and H100.
- +Drop-in replacement with zero code changes, simple API.
- −High cost with credit system and enterprise pricing, not for small teams.
- −Requires advanced skill level and dedicated infrastructure setup.
- −Numerical correctness verification is manual, potentially slow.
- −Confusing name overlaps with unrelated Forge projects.
- −Scarce independent community validation or benchmarks.
- • Credit system (1 credit/kernel, 1-2 for HF models) may require frequent purchases
- • Enterprise tier likely requires negotiated contract and significant upfront cost
Viability Score
How well maintained and how widely used is Forge CLI? Built from what the vendor actually publishes (docs, changelog, tutorials, integrations, pricing), whether the site is live, and how much real users discuss it. How we calculate this
Last calculated: October 2026
How we score →Key Features
- Generate drop-in CUDA C++ or Triton Python kernels from PyTorch modules and HuggingFace model IDs
- Swarm of 32 parallel Coder+Judge agent pairs generating and validating kernels concurrently
- MAP-Elites evolutionary optimizer with 36 behavior cells across four islands: memory_bound, compute_bound, fused_ops,
- Pattern RAG grounded in 1,711 CUTLASS patterns and 113 Triton patterns
- Reported speedups over torch.compile: Llama-3.1-8B 5.2x, Qwen2.5-7B 4.2x, Mistral-7B 3.4x, SDXL UNet 2.9x
- 97.6% correctness rate via tiered Dedup, Compile, Test and Benchmark pipeline
- Three optimization modes: --turbo (~2 min), default balanced, --quality maximum
- Kernel support across CUDA, Triton, Mojo, PyTorch and Numba
- Automatic Tensor Core optimization using WMMA and TMA on Hopper GPUs
- Input from 250+ KernelBench tasks or your own PyTorch file
- Interactive CLI wizard plus 'forge browse' KernelBench task browser
- Session management with 'forge session list' for tracking past optimizations
- Terminal-native workflow with no GUI required
About Forge CLI
Forge CLI is a terminal-first GPU kernel generation tool from RightNow AI, the YC-backed GPU research lab behind RunInfra. You point it at a HuggingFace model ID, one of 250+ KernelBench tasks, or a PyTorch file, and it writes a replacement kernel in Triton (Python) or native CUDA C++ that drops into your inference stack without a model-serving rewrite. It is built for ML platform and infrastructure engineers on datacenter NVIDIA cards — B200, H200, H100, A100 or L40S — who are paying for GPU hours they aren't using. Generation runs as 32 parallel Coder+Judge agent pairs under a MAP-Elites optimizer that spreads the search across four islands (memory_bound, compute_bound, fused_ops, tensor_cores) and 36 behavior cells. A pattern RAG layer grounds each candidate in 1,711 CUTLASS patterns and 113 Triton patterns, and every candidate runs the tiered Dedup, Compile, Test, Benchmark pipeline that produces RightNow's 97.6% correctness figure. RightNow reports 5.2x over torch.compile on Llama-3.1-8B, 4.2x on Qwen2.5-7B, 3.4x on Mistral-7B and 2.9x on SDXL UNet. Three modes set the tradeoff: --turbo finishes in about two minutes, the default is balanced, and --quality pushes maximum optimization. Automatic Tensor Core optimization covers WMMA and TMA on Hopper, and generated kernels span CUDA, Triton, Mojo, PyTorch and Numba targets. It runs entirely in a terminal, with an interactive wizard, a 'forge browse' KernelBench browser, and session tracking for past optimizations. Against GUI-driven tuners and hand-written kernel work, Forge trades interactivity for search breadth: it batch-runs many candidates rather than letting you nudge one by hand. That makes it a fit for teams optimizing at the model level, not for laptop or RTX experimentation.
Behind the Verdict
Forge is the one to reach for when torch.compile has plateaued and you still have headroom on H100 or B200 capacity. Kernel-level wins compound: a 5.2x on Llama-3.1-8B over torch.compile, if it holds on your checkpoint, changes how many GPUs you rent for the same tokens. We'd run it on the model that dominates your bill first, not on everything. The swarm design is the interesting part. Thirty-two Coder+Judge pairs under a MAP-Elites search means you're buying breadth — hundreds of candidate kernels screened through Dedup, Compile, Test, Benchmark — rather than one engineer's best attempt. The pattern RAG over 1,711 CUTLASS and 113 Triton patterns is what keeps that breadth from producing plausible garbage. Where it bites: this is a batch tool. --turbo at roughly two minutes is the floor, and --quality is not a coffee-break run. If your workflow needs a human in the loop nudging the kernel every thirty seconds, or a GUI to inspect candidates, Forge fights you. Pass if you're on consumer or RTX hardware. Nothing here targets it. Pass too if torch.compile already clears your throughput target — you'd be paying credits for headroom you don't need. Closest alternative is hand-rolled kernel work with CUTLASS or Triton and a benchmark harness, which costs senior engineer weeks per kernel. Forge compresses that, but the 97.6% correctness number means roughly one in forty candidates fails acceptance, so you still need your own numerical validation before shipping anything to production. One caveat on freshness. The vendor's self-serve pricing page publishes Starter, Pro and Team coding plans and per-token Model APIs, not Forge kernel credits — so budget from your own usage, and treat any figure not in writing as provisional. And since September 2026 the changelog has run
Researching Forge CLI? Get your full AI stack in 60 seconds.
Free, no signup — tell us your goal and get tools matched to your budget & existing stack.
Real-world workflow fit
Concrete scenarios for the personas Forge CLI actually fits — and what changes day-one when you adopt it.
You point Forge at a HuggingFace Llama-3.1-8B model ID, run the default balanced mode, and let the 32 Coder+Judge pairs submit candidates through Dedup, Compile, Test and Benchmark. Roughly a quarter of a working day later you have a Triton or CUDA C++ kernel that drops in as a PyTorch replacement.
Outcome: You benchmark the returned kernel against your existing torch.compile path and, per RightNow's published figure, may see up to 5.2x on Llama-3.1-8B — measured on your own checkpoint, not theirs.
Your Qwen2.5-7B serving path is on torch.compile and utilization is below what you provisioned. You run Forge in --turbo mode first for a fast ~2-minute pass, then re-run the promising shapes in --quality for maximum optimization, and compare the 4.2x reported figure against your own A/B.
Outcome: You either ship the faster kernel to the serving cluster or conclude torch.compile was already close enough — either answer is worth the run before committing another GPU quarter.
You use 'forge browse' to pick from 250+ KernelBench tasks, run them through the swarm, and inspect candidates with 'forge session list' to track what was generated and how each scored.
Outcome: You get a repeatable, correctness-checked kernel baseline across a broad task library instead of hand-writing each comparison.
Use Cases
- Optimize Llama-3.1-8B inference on H100 GPUs for a reported 5.2x speedup over torch.compile
- Rewrite a Qwen2.5-7B serving path to recover a reported 4.2x throughput without changing model-serving code
- Generate production-ready Triton or CUDA kernels for a HuggingFace model ID without hand-writing them
- Benchmark and compare kernel candidates automatically via the 32-pair swarm agent system
- Deploy drop-in replacement kernels with numerical correctness verified by the Dedup/Compile/Test/Benchmark pipeline
- Close the gap between provisioned GPU capacity and actual utilization on A100/H100 clusters
- Serve an SDXL UNet pipeline with a reported 2.9x faster kernel than torch.compile produces
- Run KernelBench-style evaluation across 250+ tasks using the built-in forge browse library
Models Under the Hood
as of 2026-10-17
Limitations
- The pages fetched this run document the broader RightNow site — RightNow Agent, RunInfra Model APIs and coding plans, and the changelog — and do not describe Forge CLI's own hardware requirements, generation-time behavior or pricing.
- The changelog checkpoint closes the editor line at Editor 1.0.0 with only Forge 0.1.0 referenced, and every entry after that date is RightNow Agent, so Forge's ongoing release cadence is unconfirmed by this evidence.
- Correctness is published at 97.6%, so roughly one in forty candidates fails the tiered Dedup/Compile/Test/Benchmark pipeline and needs human review.
- Forge is a batch optimization workflow you run and review yourself rather than an in-the-loop optimizer.
as of 2026-10-09
Verification history
We have re-verified Forge CLI 9 times since . Each pass re-reads the vendor's own pages and re-checks every listed field against that evidence; passes where nothing had changed are marked as such.
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
Showing the 6 most recent of 9 verification passes.
Free to cite with attribution — this page re-verifies continuously.
12-month cost
Project the real annual outlay, including the implied monthly cost when only an annual tier is published.
Vendor list price only. Add-on usage, seat overages, and contract minimums are surfaced under Hidden costs & gotchas.
Plans compared
For each published Forge CLI tier: who it actually fits, and what it adds vs. the previous tier. Cross-reference the cost calculator above for projected annual outlay.
Pay as you go
From $5 top-up
Ideal for
Apps and API workloads that bill per token rather than per seat, including teams testing hosted open models before committing to a plan.
What this tier adds
Alternative to a subscription: top up from $5, pay per token on one prepaid balance, credits valid one year, with cached input discounts on supported models.
Starter
$10/mo
Ideal for
Individual engineer trying RightNow's coding-agent plan alongside a first Forge evaluation, with modest daily usage.
What this tier adds
Starting paid tier: $10/mo with about $12 of usage, a $1.11 5-hour limit, a $2.77 weekly limit, and free standby up to $0.20 of chat a day.
Pro
$29/mo
Ideal for
Full-time engineer driving agent sessions most days and needing more headroom per 5-hour window than Starter allows.
What this tier adds
Raises the 5-hour limit to $3.23 and the weekly limit to $8.08, lifts free standby to $0.58 of chat a day, and allows 5 to 32 concurrent requests per model when busy.
Team
$99/mo
Ideal for
Small platform or infra team sharing a workspace across several engineers doing sustained kernel and agent work.
What this tier adds
Highest coding tier at $99/mo: about $119 of usage, a $10.98 5-hour limit, a $27.46 weekly limit, $1.98 daily free standby and 8 to 32 concurrent requests per model.
Enterprise
Custom
Ideal for
Organizations needing private infrastructure, custom volume, contractual SLAs and compliance attestation before running production traffic.
What this tier adds
Adds custom volume and contract terms, audit logs and RBAC, SOC 2 Type II attestation, SLAs up to 99.99%, a dedicated CSM and a private Slack channel.
Where the pricing makes sense
The company stage and team size where Forge CLI's pricing actually pencils out — and where peers do it cheaper.
Forge sits in a different buying motion from RightNow's published self-serve tiers. Those are RunInfra coding plans — $10/mo Starter, $29/mo Pro, $99/mo Team — plus pay-as-you-go Model APIs from a $5 top-up, with Enterprise on custom terms including SOC 2 Type II and SLAs up to 99.99%. Forge kernel generation runs on its own credit system (1 credit per KernelBench or custom kernel, 1-2 for HuggingFace models), and no Forge credit rate is published on the pages we reached. The economics only
Setup time & first value
How long it actually takes to get something useful out of Forge CLI — broken out by persona, not the marketing-page minute.
Terminal-native install via npm, npx, curl (macOS/Linux) or PowerShell (Windows), so getting the binary in place is minutes. First value depends on the run mode: --turbo finishes a kernel in about two minutes, the default balanced mode takes longer, and --quality can run substantially longer per model. Budget the larger time block for validating the returned kernel against your own serving path
Switching to or from Forge CLI
How to bring data in from common predecessors and how to get it back out — written for the switcher, not the buyer.
- →From hand-written CUDA C++: point Forge at your PyTorch module or kernel source and get a Triton or CUDA candidate to A/B against your current implementation.
- →From torch.compile alone: keep your existing serving code and swap the generated kernel in as a drop-in PyTorch replacement, then compare throughput on your checkpoint.
- →From a tuned vLLM deployment: use Forge to generate kernels for the specific model shapes vLLM is serving, rather than replacing vLLM itself.
- →From KernelBench scripts: use 'forge browse' to run the 250+ task library through the 32-pair swarm instead of maintaining your own harness.
- ↗To torch.compile: if the generated kernel doesn't beat your compile baseline on your own checkpoint, revert to the compiled path with no serving-code change.
- ↗To hand-written CUDA: take the Forge output as a starting point and refine it by hand if you need something the generated candidate doesn't cover.
- ↗To a tuned vLLM setup: if the bottleneck is scheduling or batching rather than kernel math, vLLM tuning may move the number more than a new kernel.
Integrations
Resources & Guides
- Documentationrightnowai.co
Docs · Forge CLI
Full product docs from rightnowai.co
- Documentationrightnowai.co
Install · Forge CLI
Full product docs from rightnowai.co
- Quickstartrightnowai.co
Quickstart · Forge CLI
Get up and running fast from rightnowai.co
- Documentationrightnowai.co
Cli Reference · Forge CLI
Full product docs from rightnowai.co
- Documentationrightnowai.co
Troubleshooting · Forge CLI
Full product docs from rightnowai.co
- Documentationrightnowai.co
Known Issues · Forge CLI
Full product docs from rightnowai.co
- Resourcerightnowai.co
Changelog · Forge CLI
Helpful link from rightnowai.co
Tutorials & Learning
Getting Started with Atlassian Forge CLI - Windows
Atlassian Developer
Getting Started with Atlassian Forge CLI - Mac OS
Atlassian Developer
What is Forge? - The Atlassian Developer Platform
Atlassian Developer
YouTube returned 6 videos for “Forge CLI”, and we withheld 1: 1 did not mention Forge CLI. Showing the 5 we can prove are about Forge CLI.
Official links
Tools that pair well with Forge CLI
Common stack mates teams adopt alongside Forge CLI, with the specific reason each pairing earns its keep.
Magic.dev
Magic builds frontier code models with ultra-long context to automate software engineering and AI research.
Deci
Automated deep learning model optimization for NVIDIA GPUs.
Claude
Claude is Anthropic's AI assistant for long-document analysis, agentic coding, and multi-step work inside one chat and Cowork surface.
Featured Head-to-Head Comparisons
Forge Cli vs Spider Cloud
Spider Cloud and Forge CLI serve completely different needs: Spider Cloud is a web data extraction API for AI agents, while Forge CLI is a GPU kernel optimizer for PyTorch models. If you need real-time web data for RAG or LLM context, Spider Cloud's freemium model and browser AI commands are the right choice. If you're an ML engineer maximizing inference speed on datacenter GPUs, Forge CLI's automated kernel generation can deliver 2–5x speedups over torch.compile, but requires contacting sales for pricing.
Forge Cli vs Temporal Ai
If you need to orchestrate reliable, fault-tolerant AI agents or microservices, Temporal AI is your pick. If your goal is maximum GPU inference speed for production models, Forge CLI delivers up to 5× faster kernels. Choose based on your bottleneck: workflow reliability vs. raw performance.
Forge Cli vs Voyage Ai
Choose Voyage AI if your priority is high-accuracy retrieval in regulated RAG workflows with long-context, domain-specific embeddings — its low-dimensional vectors and 32K token support cut storage costs and improve search. Choose Forge CLI if you need to maximize GPU inference performance for large models on datacenter hardware; recent updates show it can beat torch.compile by up to 14x with verified correctness, though it requires contacting sales for pricing and only supports enterprise GPUs.
Alternatives to Forge CLI
View all
Frequently Asked Questions
Best-of guides
Used Forge CLI? Help shape our editorial sentiment research.