Cua

What do they actually do

Cua is an open‑source framework with a paid cloud service for building “computer‑use agents” — AI systems that can see a desktop or app UI and operate it (click, type, take screenshots, run programs). Today, developers use its Python/TypeScript Agent SDK, Computer SDK, and sandboxed runtimes (local Docker, macOS VM, Windows, or Cua Cloud) to spin up isolated desktops and run agent loops against them with pluggable models YC profile, GitHub README, Docs.

You can compose different vision/LLM providers via a simple model string (e.g., a grounding model plus a planning model), benchmark agents with suites like OSWorld/HUD, and debug trajectories. The cloud product adds managed sandboxes, a web playground (Cuazar), and credit‑based billing with a free tier to try small Linux sandboxes Docs, Pricing. It’s aimed at engineers and researchers building RPA‑style automations, desktop UI tests, or data‑extraction workflows that need real GUI control rather than just APIs Docs: Use cases.

Who are their target customer(s)

• Developers/researchers prototyping agents that control desktop apps: They struggle to stitch together models, local runtimes, and secure sandboxes so an agent can actually click/type on real apps, and they need reproducible examples to iterate quickly Docs Quickstart, GitHub README.
• RPA/automation teams working with legacy desktop software: Selectors and APIs are brittle or unavailable; current scripts break with UI changes. They need a more robust way to automate full desktop workflows across OSes Docs: Use cases, YC profile.
• QA engineers building end‑to‑end tests for desktop applications: Tests are flaky across OSes and hard to reproduce. They need isolated sandboxes, benchmarking, and trajectory logs to track regressions and reliability over time GitHub: benchmarking/OSWorld, Docs.
• Data‑extraction teams pulling structured data from closed/legacy desktop apps: Manual copy/paste and fragile scrapers are slow and error‑prone. They need agents that can reliably open apps, navigate UIs, export data, and keep audit logs.
• DevOps/ML platform teams deploying agents in production: They don’t want to build/secure Windows/macOS/Linux sandboxes, manage billing, or shoulder compliance; they need managed, auditable infra with enterprise support Pricing / Enterprise.

How would they acquire their first 10, 50, and 100 customers

• First 10: Directly recruit from the OSS community (contributors, Discord, stargazers) and academic labs; hand‑hold onboarding to set up a first sandbox and seed pilots with free cloud credits GitHub, Quickstart, Pricing.
• First 50: Run hackathons and OSWorld benchmarking competitions; publish turnkey templates (RPA, testing, data extraction) and convert participants via webinars, office hours, and follow‑up credits OSWorld, Docs: Use cases.
• First 100: Add a small sales/onboarding team to run paid trials with RPA, QA, and platform teams using managed sandboxes and SLA/compliance; convert wins into case studies and channel partnerships (model/infra integrations) Pricing — Enterprise, Docs / integrations.

What is the rough total addressable market

Top-down context:

Cua sits across RPA/intelligent process automation, desktop test automation, and emerging agentic‑AI infrastructure—markets that together reach into the tens to 100+ billions by the late 2020s Gartner RPA 2024, Grand View — automation testing, MarketsandMarkets agentic AI.

Bottom-up calculation:

Use a conservative combined relevant pool of ~USD 120B from RPA + automation testing + agentic AI by the late 2020s. Then apply share scenarios for developer tools and managed sandboxes: 0.5% ≈ USD 600M, 1.5% ≈ USD 1.8B, 5% ≈ USD 6B Grand View — RPA, Grand View — automation testing, MarketsandMarkets agentic AI.

Assumptions:

• The combined pool (~USD 120B) reflects overlapping RPA, automation testing, and agentic‑AI markets in the late 2020s; Cua’s relevant slice is smaller due to focus on desktop agents.
• Adoption scenarios assume Cua sells SDKs, benchmarking, and managed sandboxes/enterprise features, not full RPA suites.
• Overlap/double‑counting across markets exists; percentages (0.5%–5%) are used to translate broad pools into Cua‑specific TAM.

Who are some of their notable competitors

• UiPath: A leading RPA platform with strong desktop automation and computer‑vision features; large enterprise footprint makes it a reference incumbent for GUI automation.
• Microsoft Power Automate (Desktop): Windows‑integrated RPA/desktop automation that targets many of the same legacy workflow use cases inside enterprises.
• Anthropic (Claude Computer Use): Model‑level ‘Computer Use’ capability that lets Claude control a virtual computer; overlaps at the agent control layer that Cua orchestrates.
• Robocorp: Developer‑focused, open‑source RPA tooling plus a managed Control Room; competes for automation developers building desktop workflows.
• SmartBear TestComplete: Commercial UI testing for desktop apps; relevant for QA teams that need reliable desktop automation and could evaluate agentic alternatives.