LaunchedEditorial Listing

Antinuous

Antinuous · Antinuous: Research Agent Swarms That Run Experiments and Simulations on Cloud Compute

Open Antinuous

Antinuous is a hosted research agent that splits a plain-language question into a swarm of agents, each with its own cloud computer, which search sources, run code, simulations and parameter sweeps in parallel and return cited sheets, notebooks, charts, decks and CAD models. It suits engineers, scientists, analysts and academics who want many experiments or source checks run at once, and its compute and tools can also be called from Claude, ChatGPT, Codex, Cursor and other agents over MCP or A2A.

PricingFreemium
Setupeasy
Runs onWeb · API
APIYes
Open sourceNo
CategoryResearch
Research AgentAgent SwarmsSimulationExperimentsParameter SweepsCFDFEACADGPU ComputeScientific ResearchCited AnswersMCPA2AModal

Best for

Engineers, scientists, quantitative analysts and academic researchers whose questions involve many independent simulations, parameter sweeps or source checks, and who want them run in parallel on hosted compute with cited, reproducible artifacts, either in the Antinuous app or from Claude, ChatGPT and coding agents

Not ideal for

Teams that need to run on their own infrastructure or keep data on premises, users who only need quick web answers, workloads that need multi-node GPU clusters today, and anyone who needs Antinuous-hosted GPUs right away rather than after 14 days of paid history

Who it's for

Engineers, scientists, academic researchers, financial and strategy analysts, go-to-market teams, and developers whose agents need parallel compute and simulation

Capabilities

  • A lead agent splits a question into smaller questions and, when the work can run in parallel, starts a swarm in which each agent takes one part
  • Work runs in rounds of plan, gather, run and review on a shared board of scored approaches, dropping weaker candidates each round, with a second model checking facts and flagging items for review
  • Each run gets its own cloud computer with simulation, engineering and data tools installed, resizable for heavy work
  • Worker pools split independent variants, seeds and configurations across many fresh workers at once (Antinuous says a sweep then takes about as long as its slowest case)
  • GPU sandboxes and jobs with up to 8 GPUs on the Max and Ultra plans, and jobs on your own Modal account on any paid plan
  • Simulation tools for CAD and meshing (STEP files and drawings), structural, frequency, buckling, heat transfer and crash analysis, CFD, circuits, batteries and multibody systems
  • Science, medicine, earth science, statistics, pricing, risk and trading models alongside Antinuous's own kernels and solvers
  • Searches the web, science databases (UniProt, PDB, AlphaFold, ChEMBL, PubChem, ClinicalTrials.gov, NCBI Gene) and public records, reads full papers and flags retracted ones, citing up to 250 sources per run
  • Opens your PDFs, spreadsheets, CAD models and data files
  • Artifacts as sheets, notebooks, documents, decks, charts, maps, tables, CAD models and protein structures, with a source or code link on each value and CSV and XLSX export
  • Decks export to PowerPoint, PDF or Google Slides
  • Drafts and other actions wait until you approve them
  • Automations that rerun research on a schedule, a sheet change or a webhook
  • MCP server for Claude, ChatGPT, Claude Code, Codex, Cursor and other MCP clients, plus A2A, to start, steer, check, stop and read runs or start swarms directly
  • Saved keys and tokens that the agent sees by name only
  • Pick the lead model for each prompt, or use Auto

Limitations

  • Hosted only: no self-hosted deployment is documented, and research requires an Antinuous account
  • No public documentation site was found; MCP setup steps are shown inside the app under Settings
  • The Free plan includes only open-weights models, no compute or sandboxes and no prepaid credits, so after the free month, experiments and simulations need a paid plan
  • Antinuous-hosted GPUs are limited to the Max ($100) and Ultra ($200) plans and open only after 14 days of paid history, with larger GPUs and more GPUs per job unlocking as paid history grows
  • Multi-node GPU clusters are currently unavailable while they are redesigned
  • Usage is metered in dollars of AI usage at provider list prices and resets weekly; past a limit you wait for the reset or, on paid plans, buy prepaid credits
  • The Free plan and free month require a verified mobile number, limited to one account per number
  • Runs take minutes rather than seconds (Antinuous says half of research runs finish within 4 minutes), so calling agents must check back
  • Speedup figures (3.5x median for swarms of 20 or more agents, 13x median for compute jobs on 16 or more workers) are Antinuous's own comparison of September to November 2026 production runs against the same work run one job at a time
  • Very young company: launched in November 2026, and Y Combinator lists a team of two
  • The specific models available on each plan are not listed by name on the pricing page

Use cases

  • Sweeping wall thickness on several bracket designs from STEP files, meshing each variant, applying load cases and comparing peak stress and deflection
  • Running thousands of vehicle, battery or production-line scenarios in parallel and comparing them with measured test data
  • Testing an investment thesis against every comparable company with a source on each figure
  • Researching every account in a market and building a sourced sheet for a sales team
  • Benchmarking kernel, code change and hardware variants across many workers
  • Letting Claude Code or Cursor hand a large experiment batch to Antinuous over MCP instead of running jobs one after another

Our take

Antinuous is aimed at a gap most research agents leave open: they read and summarize, but they do not run the hundred variants that would settle a question. Its pitch is the plumbing underneath, with a machine per run, worker pools, preinstalled solvers and cited artifacts, and the MCP and A2A access means a team already working in Claude Code or Cursor can treat it as a parallel experiment backend rather than another chat app. The pricing is clearer than most new launches, with dollar allowances at provider list prices and an optional Modal account for your own jobs. What to check before relying on it: the free plan has no compute, GPUs need the $100 plan and two weeks of paid history, multi-node clusters are paused, there is no self-hosting or public documentation, and the speedups are the company's own numbers, so trial it on a representative sweep first.

Who should use it

Mechanical, automotive and manufacturing engineers who want to sweep design variants and load cases without scripting each run; scientists and academics who need simulations and literature checks with cited, rerunnable notebooks; analysts testing a thesis across many companies with a source on each figure; and developers whose coding agents need a place to run thousands of jobs at once.

Who should skip it

Organizations that must keep data and compute on their own infrastructure, users who only need fast web answers, teams that need multi-node GPU training clusters or immediate hosted GPU access, and buyers who want a long-established vendor with public documentation and enterprise controls.

Strengths

  • Executes experiments, simulations and sweeps as code on dedicated machines rather than only summarizing literature
  • Broad built-in simulation coverage, from FEA, crash and CFD to battery, chemistry and finance models
  • Every value in an artifact links to its source or the code that produced it, and computed values keep their method and inputs
  • The same compute and tools are open to Claude, ChatGPT, Codex, Cursor and A2A agents, so it can sit underneath an existing agent workflow
  • Low entry price, a free month at Pro limits, overage billed at provider list prices with a spending cap, and the option to run jobs on your own Modal account
  • Drafts and other actions wait for approval

Weaknesses

  • Hosted only, with no self-hosting and no public documentation site
  • Free plan has no compute, so the core experiment workflow starts at $20 a month
  • Antinuous-hosted GPUs are gated behind the Max plan and 14 days of paid history, and multi-node clusters are withdrawn for now
  • Usage-based allowances that reset weekly can be hard to predict for heavy simulation work
  • Very new product from a two-person team, and its speedup figures are self-reported

Antinuous pricing

Free

$0

Billed monthly

  • $20 of AI usage a month, released weekly
  • Open-weights models only
  • No compute or GPUs
  • Cited research, agent swarms, automations, artifacts and MCP/A2A access

Pro

$20

Billed monthly

  • $100 of AI usage a month, released weekly
  • All models
  • Sandboxes and CPU jobs
  • Jobs on your own Modal account
  • No Antinuous-hosted GPUs
  • Prepaid credits past a limit, with a monthly spending cap
  • First month free at Pro limits with a verified mobile number and no card

Max

$100

Billed monthly

  • 5x Pro: $500 of AI usage a month, released weekly
  • All models
  • Sandboxes and CPU jobs
  • GPU sandboxes and jobs with up to 8 GPUs

Ultra

$200

Billed monthly

  • 20x Pro: $2,000 of AI usage a month, released weekly
  • All models
  • Sandboxes and CPU jobs
  • GPU sandboxes and jobs with up to 8 GPUs

Free tier limits: $20 of AI usage a month, released weekly, with open-weights models only, no sandboxes, CPU jobs or GPUs, and no prepaid credits past the limit. Requires a verified mobile number, one account per number.

Note: AI usage (model inference and web retrieval) is measured at provider list prices; compute covers the sandboxes and jobs agents run. Both reset weekly. Past a limit you can wait for the reset or, on paid plans, use prepaid credits billed at provider list prices with no markup, with automatic recharge up to a monthly spending cap you set. GPUs open on Max and Ultra after 14 days of paid history, and more or larger GPUs per job unlock as paid history grows. Paid plans can also run jobs on your own Modal account, which do not count toward your plan. Upgrades apply once paid; downgrades and cancellations apply at the end of the billing period. The free month needs no card and excludes GPUs; afterwards you stay on Free until you choose a plan.

Technical specs

Modalities

Text, Documents, Spreadsheets, CAD (STEP), Code, Charts, Slides

API pricing

Included in plans; MCP and A2A calls use the same weekly AI usage and compute allowances as the app

Where Antinuous excels

Design variant studies for mechanical parts

Agents can open STEP files, mesh every variant, apply the load cases and compare stress and deflection across workers in parallel, then return a notebook with the code and settings for each run.

Handing batch experiments off from a coding agent

Claude Code, Codex or Cursor can start a swarm over MCP, keep working while it runs, and read the results with the evidence behind each value instead of running jobs one at a time locally.

Sourced comparison research

Swarms split a market or company list across agents and return a sheet in which each figure links to its source, exportable to CSV or XLSX, with scheduled reruns to keep it current.

Antinuous vs. competitors

Antinuous vs. ClawsGO Science

ClawsGO Science gives one agent its own cloud Linux host to work for hours on literature reviews, analysis and LaTeX manuscripts. Antinuous splits a question across a swarm of agents, each on its own machine, and leans on parallel simulation and parameter sweeps, returning sheets, notebooks, decks and CAD models instead of a paper draft.

Antinuous vs. OpenScience

OpenScience is an open-source (Apache 2.0) research agent you run with your own model and compute, keeping work on your own files. Antinuous is hosted and proprietary, but supplies the compute, worker pools, GPU sandboxes and preinstalled simulation tools for you.

Antinuous vs. Prism by Navier AI

Prism is a desktop engineering agent that runs CAD, FEA and CFD with open-source solvers on your own machine. Antinuous runs engineering simulations in the cloud across many workers at once, and also covers science, finance and market research.

Antinuous vs. The AI Scientist

The AI Scientist is Sakana AI's open-source pipeline that drives coding agents through ML experiments to a LaTeX paper on your own hardware. Antinuous is a hosted platform for parallel experiments across engineering, science and finance, with cited artifacts rather than a paper as the output.

Frequently asked questions

What does Antinuous do?

Antinuous is a hosted research agent and compute platform. You ask a question in plain words; a lead agent splits it into smaller questions and, when the work can run in parallel, starts a swarm of agents that each get their own computer to search sources, run code, simulations and sweeps. Results come back as sheets, notebooks, documents, decks, charts or CAD models, with each value linked to its source or code.

How much does Antinuous cost?

Free is $0 with $20 of AI usage a month and open-weights models only, without compute. Pro is $20 a month with $100 of AI usage, all models, sandboxes and CPU jobs. Max is $100 a month (5x Pro) and Ultra is $200 a month (20x Pro), both adding GPU sandboxes and jobs with up to 8 GPUs. Allowances reset weekly, paid plans can buy prepaid credits past a limit, and the first month is free at Pro limits with a verified mobile number and no card.

Can I use Antinuous from Claude, ChatGPT, Codex or Cursor?

Yes. Antinuous runs an MCP server that Claude, ChatGPT, Claude Code, Codex, Cursor and other MCP clients can connect to, and agents that speak A2A can use it too. An outside agent can start Antinuous's agent, send follow-ups, check or stop a run, read results and the evidence behind each value, start a swarm directly, or change results. It signs in with your Antinuous account and sees only what you can see.

Does Antinuous include GPUs?

Only on the Max and Ultra plans, and only after 14 days of paid history. Larger GPUs and more GPUs per job, up to 8, open as paid history grows. Multi-node GPU clusters are being redesigned and are not currently available. On any paid plan you can also save a Modal token in Settings and run jobs on your own Modal account, which do not count toward your plan.

What kinds of simulations can Antinuous run?

Antinuous lists CAD and meshing with STEP files, static, frequency, buckling, heat transfer and crash analysis, CFD and airfoil screening, circuit, battery, multibody, hydraulic and thermal system models, plus life science, chemistry, earth science, statistics, pricing, risk and trading models. They run as code in each run's own computer or across worker pools.

Is Antinuous open source or self-hostable?

No. Antinuous is a hosted service, and no open-source release or self-hosted deployment is documented. The closest option to bringing your own infrastructure is running jobs on your own Modal account on a paid plan.

Is Antinuous still in beta?

No. Antinuous ran a beta for a few weeks and launched publicly on November 19, 2026, with open sign-up and paid plans. It is part of Y Combinator's Spring 2026 batch, and Y Combinator lists a team of two.

Integrations & fit

ClaudeChatGPTClaude CodeOpenAI CodexCursorMCPA2AModalSlackGoogle SlidesPowerPointGitHubarXivSEC EDGARFREDBLSUniProtPDBAlphaFoldChEMBLPubChemClinicalTrials.govNCBI Gene
Good fit forSolo / individual, Startup / small team
Pricing modelFreemium· Free tier available
See pricing on Antinuous →

Alternatives to consider

About Antinuous

Antinuous, a two-person San Francisco startup in Y Combinator's Spring 2026 batch, launched publicly on November 19, 2026 after a few weeks of beta. It describes itself as infrastructure for autonomous research, experiments and simulations, built in four layers: elastic compute, swarm orchestration, simulation, and artifacts. You type a question or send it from a coding agent. A lead agent breaks it into smaller questions, and when the work can be parallelized it starts a swarm in which each agent takes one part. Runs work in rounds (plan, gather evidence, run simulations, review) on a shared board of scored approaches; weaker candidates are dropped after each review, so later rounds run fewer jobs, and a second model checks facts and flags the ones that need your review. On paid plans, every experiment runs as code on its own computer with simulation, engineering and data tools installed, and large jobs fan out across pools of workers. GPU jobs run on Antinuous clusters on the Max and Ultra plans, and any paid plan can send jobs to your own Modal account instead. The simulation layer covers CAD and meshing, finite element structures and crash, CFD, circuits and batteries, life science and chemistry models, and pricing and risk models, while the research side searches the web, science databases such as UniProt, PDB, ChEMBL and ClinicalTrials.gov, and public records, reads full papers and flags retracted ones. Results come back as sheets, notebooks, documents, decks, charts, maps, CAD models with STEP files or protein structures, with each value linked to its source or the code that produced it, and drafts and other actions wait for your approval. Unlike literature-focused research agents, it runs the experiments as code, and it exposes the same tools to outside agents over MCP and A2A. The tradeoffs: it is a very young, hosted-only service with no public documentation site, usage is metered in dollars of AI usage and compute that reset weekly, Antinuous-hosted GPUs require the $100 plan plus 14 days of paid history, multi-node GPU clusters are withdrawn for a redesign, and the speedups quoted on its site (3.5x for swarms, 13x for compute jobs, both medians) are the company's own measurements of its production runs.

Updates from Antinuous

LaunchAntinuous launches publicly

After a few weeks of beta, Antinuous (Y Combinator Spring 2026) launched its research and experimentation platform, combining swarm orchestration, compute clusters and simulation software, with Free, Pro, Max and Ultra plans and MCP and A2A access for outside agents.

Are you the founder? Claim this listing →