Cua

Infrastructure

medium confidence

cua.ai ↗ · Status: active confirmed · Founded 2025

Cua (trycua, YC X25) builds open-core 'computer-use 2.0' infrastructure: an MIT, agent-neutral background driver for macOS, Windows and Linux (CLI/MCP), local and cloud sandboxes, elastic Cua Fleets for parallel eval and RL training (cloud, BYOC, on-prem), Lume macOS virtualization, the Cua Spaces Mac app, small CUA-S1 decision models, and Cua-Bench for verifiable tasks, leaderboards and RL trajectory export. SOC 2 Type I.

Key facts

Headquarters

San Francisco, CA, USAreported cite

Headcount band

1-10confirmed cite

Total raised

$500Kreported cite

Last round

Pre-seed/seed, ~$500K, ~June 2025reported cite

SOC 2

Type IIconfirmed cite

What they sell

infraconfirmed cite

Open source

yesconfirmed cite

Deployment

Local open-source sandboxes (GUI Docker/QEMU/Apple VZ, Lume) + managed Cua Fleet cloud (run.cua.ai) + BYOC + on-prem + SDK/CLI/MCPconfirmed cite

What's new

12 sourced updates · last 6 months

1 Oct 2026product Cua launches Cua Spaces v0.1.0, a source-available Mac app for watching and supervising agent desktops
Cua Spaces is 'the Mac app for working alongside your agents. Watch, take over, share files and teleport apps' (cua.ai/docs). It ships as a signed .dmg and .pkg for macOS 26+. It is licensed FSL-1.1-MIT (source-available, becomes MIT two years after release), and hosted or managed-service use needs commercial terms. Fast follow-up releases reached v0.7.2 by 2026-10-05.
Why it matters: First non-MIT (source-available) component in the stack, which matters for commercial-use review.
25 Sep 2026product Cua Driver adds compositor-native background computer use on Hyprland (Omarchy)
A compositor-native synthetic cursor lets an agent operate a background window on Hyprland while the user keeps their own pointer. As of 2026-10-05 the latest cua-driver-rs release is v0.33.4, which adds multi-monitor Hyprland support.
Why it matters: Extends Linux coverage beyond X11 to a Wayland compositor.
18 Sep 2026open-source Cua releases CUA-S1, small open 'System One' models for narrow computer-use decisions
Dillon DuPont and Francesco Bonacci introduced CUA-S1-FORMS, an MIT-licensed 706k-parameter model (2.8 MB checkpoint). Instead of generating tokens, it scores the available actions on form fields (use, check, click, skip). Show HN on 2026-09-19 got 95 points. Cua's accuracy and latency comparison against hosted Jev is self-reported.
Why it matters: Cua is now shipping small specialised models alongside its infrastructure: cheap local decisions inside larger agent loops.
11 Aug 2026research Lume Metal capability shim speeds llama.cpp inference inside macOS VMs by 11-16x
Francesco Bonacci and Johnny Franks built an open-source, process-scoped Metal capability shim (libs/lume/metal-capability-shim) that lets llama.cpp in macOS guests choose optimised GPU kernels. On an M1 Ultra with TinyLlama 1.1B: 11.08x faster prompt processing and 16.36x faster generation. It got about 307 points on Hacker News.
Why it matters: Makes it practical to run local models inside sandboxed macOS VMs for agent eval.
6 Aug 2026product Cua Driver 0.19.0 adds extension-free browser use via Chrome DevTools Protocol
Agents can combine page-aware CDP actions with OS-level control in one session, with no Chrome extension. In a preliminary OSWorld 2.0 test on 37 valid Chrome task pairs (GPT-5.6 Sol, 80 steps), the mean score moved from 0.0043 to 0.0298 (+2.55 pp). That was not significant: 95% CI -0.0054 to +0.0766, p = 0.5.
Why it matters: Puts browser use and computer use in one environment interface. The vendor reports a non-significant result candidly.
13 Jul 2026product Cua publishes a recap of its 'Computer-Use 2.0' talk at AI Engineer World's Fair: Cua Driver 0.8, the Cua-Bench catalog and Cua Fleets
Francesco Bonacci, Dillon DuPont and Robert Wendt presented at AI Engineer World's Fair 2026 (swyx invited them in May; the post does not give the talk date). Cua Driver 0.8 was stable on macOS, Windows and Linux. Named Driver adopters: Clicky (first, in April), then Hermes, Qwen Code, H Company and Factory's Droids. H Company is described as a partner that helped audit the regression matrix. At talk time the Cua-Bench catalog had 130 verifiable tasks across 42 environments and five platforms. Cua Fleets manages warm sandbox pools for eval and RL training on Linux, Windows and Android, with macOS in preview.
Why it matters: Shows Cua moving from single sandboxes to fleet-scale RL training and eval infrastructure. The adopter list is self-reported, except Factory, whose own docs confirm the integration.
24 Jun 2026benchmark Cua evaluates Gemini 3.5 Flash computer use on Cua-Bench's KiCad suite, with early access from Google
Cua says: 'Google gave us early access to the Computer Use tool powered by Gemini 3.5 Flash through the Gemini API.' On the 25-task KiCad suite with a 200-step budget, Gemini 3.5 Flash had the highest mean reward in the post (0.267; 5 full solves, 3 partial). GPT-5.5 got 0.240 (6 full, 0 partial), Claude Sonnet 4.5 and Haiku 4.5 0.253 each, and Claude Opus 4.8 0.240.
Why it matters: Vendor-stated pre-release access from a frontier lab, plus a published cross-lab comparison of computer-use models.
18 Jun 2026product Cua Driver comes to Linux (X11/XWayland; native Wayland in preview)
Release tag 0.5.7 drives real Linux desktop apps in the background over MCP or CLI. It uses AT-SPI 2 over D-Bus for the accessibility tree and XTEST for input. Native Wayland is opt-in preview only. Tested on Debian 12, Ubuntu 22.04 and 24.04, Rocky 9 and Fedora 41.
Why it matters: Cua Driver now covers macOS, Windows and Linux behind one agent-neutral interface.
15 Jun 2026partnership Snorkel AI publishes a 25-task KiCad (EDA) benchmark built on Cua-Bench
Snorkel researchers Zhengyang (Jason) Qi and Armin Parchami wrote 25 expert schematic-editing tasks in KiCad. Grading is verifiable: the netlist is compared to a ground-truth final project state. Cua provided the harness and verifier support. In Snorkel's own analysis, the best agent tested passed 4/25 tasks (16%) and 0/16 build-from-scratch tasks.
Why it matters: A third-party RL data vendor publicly confirms the collaboration and is using Cua-Bench to build professional-software tasks.
27 May 2026product Cua Driver adds Windows background computer use with synthetic per-agent cursors
Cua Driver now supports background computer use on Windows for Claude Code, Codex, Hermes and any MCP or CLI agent loop. Each agent gets its own visible synthetic cursor, painted and recorded by the driver and kept separate from the user's physical pointer.
Why it matters: Brings the same driver to Windows desktop software for agent evaluation and data generation.
May 2026other Cua receives SOC 2 Type I report; Type II observation period under way
The trust page says: 'Cua received its SOC 2 Type I report in May 2026.' The report is available to customers and qualified prospects on request (enterprise@cua.ai). The company is 'completing the operating-effectiveness observation period' for Type II. No auditor is named.
Why it matters: Removes a procurement blocker. The baseline record had SOC 2 as unknown and no security page.
23 Apr 2026open-source Cua open-sources Cua Driver for macOS: background, multi-cursor computer use for any agent
Cua released cua-driver under a permissive licence (MIT per the repo). It lets Claude Code, Codex, other MCP clients or a custom loop drive native macOS apps in the background over CLI or MCP, using the private SkyLight framework (SLEventPostToPid) so the agent never takes the user's cursor. A Show HN on 2026-04-28 got about 190 points.
Why it matters: Cua positions background computer use as 'a commodity, not a feature of one agent's product'. Researchers get a harness-agnostic way to drive real desktop apps.

Founding team

Francesco Bonacci

Founder (YC lists Founder/CTO; cua.ai About page lists him under Product)

Ex-Microsoft (Xbox, Microsoft AI). Co-author of Microsoft's Windows Agent Arena benchmark (2024). Built Lume before founding Cua in March 2025.

Profile ↗

Dillon DuPont

Research lead, early team (joined May 2025; the About page does not call him a co-founder)

Ex-Microsoft. Co-author of Windows Agent Arena (2024). Leads Cua research (Gemini 3.5 Flash eval; co-author of CUA-S1).

Profile ↗

What practitioners should know

For researchers

Cua-Bench is MIT-licensed and supports trajectory export for RL/SFT. The KiCad EDA leaderboard (25 tasks, authored by Snorkel AI) tops out at a mean reward of about 0.27, far from saturated. Every full pass was an edit, and no model built a schematic from scratch (https://cua.ai/cuabench/leaderboard).
Cua publishes detailed engineering write-ups (macOS SkyLight, Windows synthetic cursors, Linux AT-SPI/XTEST, Hyprland) and reports null results, e.g. OSWorld 2.0 browser-tools test at +2.55 pp, not significant (p = 0.5).
Bonacci, DuPont and Dan Zhao co-authored Microsoft's Windows Agent Arena benchmark before Cua, so the team has real benchmark-building pedigree.

For program managers

Small YC X25 team. YC still lists team size 3, but cua.ai/about names 8 current members across SF, Toronto, Vancouver, New York and New Jersey. No funding beyond the ~$500K June 2025 seed (aggregator-reported: YC, Mark Kahn).
Frontier-lab ties are vendor-stated: early access to Google's Gemini 3.5 Flash Computer Use tool. The homepage lists OpenAI under 'ecosystem integrations', but the link goes to an OpenAI-staff PR in cirruslabs/macos-image-templates ('Include Metal Capabilities Helper') that does not name Cua. Treat it as an integration claim, not a customer.
Third-party-verified relationships: Snorkel AI (KiCad benchmark on Cua-Bench) and Factory (Droid Control docs install cua-driver). Hermes, Qwen Code, H Company and Clicky are self-reported adopters or partners.
Licensing is open-core: MIT for the SDK, CLI, Driver, Lume and Bench; FSL-1.1-MIT for Cua Spaces; AGPL-3.0 and Apache-2.0 for optional perception extensions; hosted commercial use needs review per COMMERCIAL.md.
Hiring (YC): Founding Technical GTM Lead; Founding Engineer, Infra & Agent Systems; Research Intern (Summer 2026).

For data ops

SOC 2 Type I report received May 2026, available on request via enterprise@cua.ai. Type II observation period in progress. No other certifications listed (https://cua.ai/trust).
Deployment: local open-source sandboxes (Docker/QEMU/Apple VZ, Lume), Cua Fleet cloud via run.cua.ai, BYOC and on-prem. The Fleet price per vCPU-hour and per GiB-hour is published on https://cua.ai/pricing.
Integration surfaces: Cua Driver over CLI and MCP; SDKs in Python, TypeScript, Swift and Rust (Kotlin bindings in the repo).

Benchmarks & research

3 published

Cua-Bench EDA Leaderboard (KiCad, v1)leaderboard · Jun 2026

25 expert-authored KiCad schematic tasks with a 200-step budget and verifiable netlist grading. The tasks were authored by Snorkel AI on Cua-Bench.

Claude Fable 5: 6/25 full + 2 partial, mean reward 0.267. GPT-5.5: 6/25 full, 0.240. Gemini 3.5 Flash: 5/25 full + 3 partial, 0.267. Claude Sonnet 4.5 and Haiku 4.5: 0.253 each. Every full pass was an edit to an existing schematic.

Open ↗Paper ↗

Cua-Benchbenchmark · Dec 2025

Cua's open (MIT) collection of desktop and mobile computer-use tasks, with a harness for evaluation and training, plus task building, agent evaluation, leaderboards and trajectory export for RL/SFT. Launched December 2025. Listed contributors include Tianbao Xie (OSWorld lead author) and Alex Shaw (Terminal-Bench co-creator). At the AIEWF 2026 talk the catalog had 130 verifiable tasks across 42 environments and five platforms.

The best frontier agent tested clears 6 of 25 expert KiCad tasks. No model reliably builds a schematic from a blank canvas.

Open ↗

Windows Agent Arena (Microsoft; co-authored by Cua's Bonacci, DuPont and Dan Zhao before Cua)paper · 12 Sep 2024

Microsoft benchmark for multimodal OS agents on Windows, with more than 150 tasks and parallelised cloud evaluation. Cua presented it at ICML 2025.

The Navi agent reached 19.5% success in the Windows domain, against 74.5% for unassisted humans.

Open ↗

Products

Cua Driver ↗

MIT open-source background computer-use driver for macOS, Windows and Linux (including Hyprland), exposed over CLI and MCP. Works with Claude Code, Codex and custom loops. Adds extension-free browser use via CDP from v0.19. Latest cua-driver-rs release is v0.33.4 (2026-10-05).launched 23 Apr 2026

Cua Sandbox / SDK ↗

Local sandboxes (GUI Docker, QEMU, Apple VZ) plus cloud sandboxes. The SDK is documented for Python, TypeScript, Swift and Rust; the repo also lists Kotlin bindings.

Cua Fleets ↗

Usage-metered elastic fleets with warm sandbox pools for parallel eval and RL training. Linux, Windows and Android, with macOS in preview. Console at run.cua.ai. BYOC and on-prem are available.launched 2026

Cua-Bench ↗

Computer-use task collection and harness: evaluation, leaderboards and trajectory export for RL/SFT.launched Dec 2025

Lume ↗

Local macOS and Linux VMs on Apple Silicon via Virtualization.framework. Includes an open-source Metal capability shim for faster in-VM inference.

Cua Spaces ↗

Source-available (FSL-1.1-MIT) Mac app for watching, taking over and sharing files with agent desktops (Spaces).launched 1 Oct 2026

CUA-S1 ↗

Small MIT-licensed 'System One' decision models for narrow UI decisions. The first release, CUA-S1-FORMS (706k parameters), handles forms.launched 18 Sep 2026

Scale & velocity

Current headcount

~8 (8 current team members named on cua.ai/about; YC profile still shows team size 3)confirmed cite

Headcount growth

unknown

Open roles

3 (Founding Technical GTM Lead; Founding Engineer, Infra & Agent Systems; Research Intern Summer 2026)confirmed cite

Other locations

unknown

Distributed / remote

unknown

Research depth

Has researchers

yesreported cite

Researcher count

unknown

Backgrounds

Francesco Bonacci, founder: ex-Microsoft (Xbox, Microsoft AI), Windows Agent Arena co-author, Dillon DuPont, research lead: ex-Microsoft, Windows Agent Arena co-author, Dan Zhao, research: Windows Agent Arena co-author, Kai Mei, research (background not detailed on the About page)confirmed cite

Capital

Total raised

~$500K (pre-seed/seed, ~June 2025)reported cite

Last round

Pre-seed/seed, ~$500K, ~June 2025reported cite

Investors

Y Combinator (X25 batch)reported cite

Valuation

unknown

Revenue signals

unknown

Security & compliance

SOC 2

SOC 2 Type I (report received May 2026, available on request); SOC 2 Type II observation period in progressconfirmed cite

Other certifications

unknown

Security page

https://cua.ai/trustconfirmed cite

Product

What they sell

infraconfirmed cite

Open source

yesconfirmed cite

License

MIT (default: SDK, CLI, Driver, Lume, Bench); Cua Spaces FSL-1.1-MIT (source-available, becomes MIT 2 years after release); optional extensions AGPL-3.0 / Apache-2.0; commercial hosting per COMMERCIAL.mdconfirmed cite

Deployment model

Local open-source sandboxes (GUI Docker/QEMU/Apple VZ, Lume) + managed Cua Fleet cloud (run.cua.ai) + BYOC + on-prem + SDK/CLI/MCPconfirmed cite

Maturity

GA: open-source components plus a usage-metered Cua Fleet. Cua Driver stable on macOS/Windows/Linux (v0.33.x by Oct 2026). Cua Spaces newly launched (v0.x, Oct 2026). Fleets macOS in preview.confirmed cite

Notable customers

Snorkel AI (KiCad benchmark built on Cua-Bench) verified Factory (Droid Control integrates cua-driver) verified H Company (partner; helped audit Driver regression matrix) self-claimed Qwen Code (Driver adopter / ecosystem integration) self-claimed Hermes (Driver adopter) self-claimed Clicky (first Driver adopter) self-claimed ⚑Google (early access to Gemini 3.5 Flash Computer Use tool; not a customer) self-claimed ⚑OpenAI (homepage 'ecosystem integration'; link is an OpenAI-staff Tart image PR that does not name Cua) self-claimed cite

Buyer analysis

Best fit: Teams that need agent-neutral real-desktop environments (macOS/Windows/Linux apps) at fleet scale to evaluate computer-use agents or generate RL/SFT trajectories, including professional-software tasks such as EDA.

How we verified this

2026-10-05 refresh: I opened every draft source directly. The release history is well supported by primary sources: Cua blog posts, the GitHub repo and releases, the trust and pricing pages, and Snorkel's own post. All dated news items fall inside the window. Corrections: - **Founders:** the draft's co-founder Alessandro Puppo is unsupported. YC and cua.ai/about name only Francesco Bonacci as founder, so Puppo was dropped. Dillon DuPont was added as research lead with a verified photo from the About page. - **Spaces:** the Lume/gVisor claims do not appear in the release notes and were removed. - **Snorkel:** the result was misattributed and was corrected to Snorkel's own figures (4/25 and 0/16 build-from-scratch). - **Cua-Bench launch date:** corrected to Dec 2025. - **Unverifiable harness/format details:** decorators, Parquet and adapters were removed. - **X links:** unreachable, so dropped. - **Customers:** Hugging Face and Meta were removed because the logo wall is gone. Factory was upgraded to verified via Factory's docs. The OpenAI 'integration' was caveated: its link is an OpenAI-staff Tart PR that does not name Cua. Additions I found: - CUA-S1 model launch (Sep 2026). - Hyprland support (Sep 2026). - The 8-person team on the About page. - The team's Windows Agent Arena paper. I found no new funding, but the WebSearch budget was exhausted mid-session, so the funding sweep used direct fetches (Google News RSS, the HN Algolia API, the aggregator page, the company timeline) rather than broad search.

Related vendors

Sources

cua.ai/ · 2026-10-05, Homepage: tagline, products, ecosystem integrations (OpenAI link -> cirruslabs PR #372), SOC 2 Type I, BYOC/on-prem, 28.1K; no logo wall
cua.ai/about · 2026-10-05, Story (Francesco founded Cua; Dillon joined shortly after), Journey timeline, 8 named team members, founder photos
cua.ai/pricing · 2026-10-05, Single usage-based Cua Fleet plan; local/cloud/on-prem; BYOC; SOC 2 Type I
cua.ai/trust · 2026-10-05, SOC 2 Type I report received May 2026; Type II observation period
cua.ai/blog · 2026-10-05, Blog index with 2026 post dates and authors
cua.ai/blog/inside-macos-window-internals · 2026-10-05, 2026-04-23 Cua Driver open-sourced for macOS
cua.ai/blog/inside-windows-computer-use · 2026-10-05, 2026-05-27 Windows Driver, synthetic cursors
cua.ai/blog/inside-linux-computer-use · 2026-10-05, 2026-06-18 Linux Driver 0.5.7; tested distros
cua.ai/blog/evaluating-gemini-3.5-flash-on-computer-use · 2026-10-05, 2026-06-24 Gemini 3.5 Flash eval; Google early access; per-model results
cua.ai/blog/computer-use-2-ai-engineer-worlds-fair · 2026-10-05, 2026-07-13 AIEWF recap: Driver 0.8, adopters, H Company, Cua-Bench 130 tasks/42 envs, Fleets
cua.ai/blog/extension-free-browser-use · 2026-10-05, 2026-08-06 Driver 0.19.0 CDP browser use; OSWorld 2.0 n.s. result
cua.ai/blog/gpu-passthrough-macos-vms · 2026-10-05, 2026-08-11 Lume Metal capability shim
github.com/trycua/cua/blob/main/blog/omarchy-cua-driver.md · 2026-10-05, 2026-09-25 Hyprland/Omarchy compositor-native background computer use
news.ycombinator.com/item?id=49767564 · 2026-10-05, Show HN: CUA-S1 (2026-09-19), CUA-S1-FORMS 706k params, MIT
github.com/trycua/cua · 2026-10-05, Repo: 28.1k stars, components, licence table
github.com/trycua/cua/releases · 2026-10-05, cua-driver-rs v0.33.4, cua-spaces v0.6.0-v0.7.2 (Oct 3-5)
github.com/trycua/cua/releases/tag/cua-spaces-v0.1.0 · 2026-10-05, Cua Spaces v0.1.0 released 2026-10-01
github.com/trycua/cua/tree/main/apps/cua-spaces · 2026-10-05, Cua Spaces README: FSL-1.1-MIT
cua.ai/docs · 2026-10-05, Docs: SDK (Python/TS/Swift/Rust), Driver, Spaces ('The Mac app...'), Bench, Lume, CLI
cua.ai/cuabench · 2026-10-05, Cua-Bench page; contributors; Snorkel x Cua report
cua.ai/cuabench/leaderboard · 2026-10-05, EDA KiCad v1 leaderboard results
snorkel.ai/cua-bench-computer-use-agent-benchmark/ · 2026-10-05, Snorkel 2026-06-15 report on the KiCad tasks built on Cua-Bench
docs.factory.com/cli/features/droid-control · 2026-10-05, Factory docs install and use cua-driver (third-party confirmation)
github.com/cirruslabs/macos-image-templates/pull/372 · 2026-10-05, Target of homepage 'OpenAI' integration link; OpenAI-staff PR, merged 2026-08-19, does not name Cua
www.ycombinator.com/companies/cua · 2026-10-05, YC profile: Bonacci Founder/CTO + photo, team size 3, 3 jobs
github.com/francedot · 2026-10-05, Francesco Bonacci GitHub (pins trycua/cua, microsoft/WindowsAgentArena)
arxiv.org/abs/2409.08264 · 2026-10-05, Windows Agent Arena paper; authors include Bonacci, DuPont, Zhao
startupintros.com/orgs/cua · 2026-10-05, Aggregator: $500K seed 2025-06-01, YC + Mark Kahn (reported only); no newer round
codenote.net/en/posts/codex-computer-use-cua-repl-vs-trycua-cua-driver/ · 2026-10-05, 2026-10-01 teardown: OpenAI cua_repl shares no code with trycua cua-driver (name-collision check)
www.linkedin.com/company/cua-ai · 2026-06-07, Public snippet: founded 2025, SF, 2-10 employees band, pre-seed $500K July 2025
tracxn.com/d/companies/cua-ai · 2026-06-07, Funding $500K seed/pre-seed, founders Francesco Bonacci and Alessandro Puppo, founded 2025
cuabench.ai/ · 2026-06-07, Cua-Bench: runs OSWorld, ScreenSpot, Windows Arena; parallel execution, trajectory export for RL, dataset versioning
news.ycombinator.com/item?id=43773563 · 2026-06-07, Launch HN: Cua (YC X25) Open-Source Docker Container for Computer-Use Agents
www.linkedin.com/in/francesco-bonacci-70428a121/ · 2026-06-07, Co-founder Francesco Bonacci
www.linkedin.com/in/alessandro-puppo/ · 2026-06-07, Co-founder Alessandro Puppo

Last updated 2026-10-05 · Every quantitative field carries a source and a confidence tag. Fields we could not source publicly are marked unknown, never estimated. See the methodology.