Anthropic Acquires Vercept to Advance Computer Use Automation
On February 25, 2026, Anthropic announced the acquisition of Vercept, a Seattle-based startup specialized in computer-use automation. The nine-person team, previously building Vy (a cloud-based agent for complex multi-step desktop automation), will join Anthropic to accelerate Claude's computer use capabilities. This acquisition signals Anthropic's commitment to moving beyond text-based interactions toward genuinely autonomous agents capable of navigating complex graphical interfaces. For OpenClaw users and operators, understanding what Vercept brings and what it means for the roadmap is essential.
Vercept's Expertise and Vy Platform
Vercept was founded to solve a specific problem: while large language models are excellent at reasoning about instructions, they historically struggled with precise, reliable interaction with graphical user interfaces. The company's Vy platform operated as a cloud-based agent that could open applications, navigate menus, click buttons, fill forms, and perform complex desktop workflows—all while maintaining context about what it was trying to accomplish.
Vy handled real-world complexity: waiting for applications to load, adapting to UI variations across operating systems and versions, recovering from transient failures, and maintaining state across multi-step operations. These are problems that seem simple in abstraction but prove devilishly difficult in practice.
Previous Funding and External Shutdown
Vercept raised $50 million in venture funding, indicating significant investor confidence in the computer-use automation market opportunity. However, with the acquisition by Anthropic, Vercept's external products are being shut down effective March 25, 2026. This is typical in acqui-hires: Anthropic is interested in the team and their expertise, not necessarily in continuing to operate Vy as a standalone product.
Strategic Fit with Claude
The acquisition accelerates Claude's multi-step task automation capabilities. While Claude already supports computer use (vision input to understand UI state, planning and tool usage for actions), Vercept's expertise will deepen this capability:
- Reliability at scale: Handling failures, timeouts, and edge cases that occur in real production environments
- Complex multi-step workflows: Orchestrating sequences of actions across multiple applications
- Adaptation and recovery: Detecting when expected UI elements are absent and adapting gracefully
- Performance optimization: Minimizing the time and computational resources required for automation tasks
Current State of Claude Computer Use
Claude's computer use capabilities, available via extended beta, already support impressive automation. The model can:
- Take screenshots to understand current UI state
- Move the cursor and click elements
- Type text and interact with forms
- Chain actions across multiple steps and applications
Performance on OSWorld benchmarks (a standard for measuring computer use automation) improved dramatically between late 2024 and early 2026, reaching 72.5% with Sonnet 4.6. This indicates the capability is already quite capable, but there's clearly room for improvement on complex, real-world tasks.
What Improving Computer Use Means for OpenClaw
For OpenClaw deployments, better computer use automation opens new use cases:
Desktop Application Automation: Agents that can interact with legacy desktop applications, ERP systems, and specialized industry software. Many enterprises still rely on Windows-based applications that resist API-driven automation. OpenClaw agents with reliable computer use become force multipliers for these organizations.
Browser Automation at Scale: Web scraping, form filling, and multi-step web workflows become more reliable and maintainable than traditional Selenium or Playwright-based automation scripts.
RPA Replacement: Robotic Process Automation (RPA) platforms like UiPath have been dominant in enterprise automation, but they're expensive and rigid. OpenClaw agents with reliable computer use become competitive alternatives: more flexible, cheaper, and capable of handling variation that traditional RPA struggles with.
Cross-System Orchestration: Agents that can navigate multiple systems—legacy internal software, SaaS applications, cloud services—to orchestrate complex business processes that don't have good API bridges.
Integration with OpenClaw's Tool Policy System
OpenClaw's permission model is crucial when enabling computer use. With Vercept's improvements, computer use becomes more powerful and more dangerous: an agent with unrestricted screen interaction capabilities can do real damage (delete files, modify data, access sensitive information).
OpenClaw's tool policies should restrict computer use explicitly:
- Limit which applications agents can interact with (block access to password managers, personal data, security settings)
- Require explicit approval before certain irreversible actions (deleting files, modifying system settings)
- Log all computer use interactions for audit purposes
- Sandbox agents so computer use is isolated to specific environments, not production systems
- Implement timeouts and resource limits to prevent runaway automation
Security Considerations
Improved computer use capability also expands the attack surface. An agent with reliable ability to interact with graphical interfaces could potentially:
- Bypass application-level access controls by manipulating the UI
- Extract sensitive data visible on screen (PII, credentials, confidential documents)
- Interact with unintended applications if tool policies are too loose
- Generate screenshots containing sensitive information, which then get stored in logs
Mitigating these risks requires careful design:
- Run computer use agents in isolated environments (virtual machines, containers) with no access to sensitive data or critical systems
- Implement strict tool policies limiting which applications agents can interact with
- Redact sensitive information from screenshots before storing them
- Require human approval before agents execute particularly risky operations
- Monitor computer use for anomalous patterns that suggest compromise or misuse
OpenClaw Computer Use Roadmap
With Vercept joining Anthropic, expect significant improvements to computer use over the coming quarters. Likely enhancements:
- Better UI understanding: Improved ability to interpret complex, dynamic interfaces
- Faster execution: More efficient action sequences reducing the number of API calls required
- Cross-OS support: Current computer use is web-focused; look for expansion to Windows, macOS, and Linux desktop automation
- Application-specific optimizations: Better handling of specific application categories (ERP systems, CRM platforms, etc.)
- Better error handling: More graceful recovery from unexpected UI states, timeouts, and failures
When to Use Computer Use in OpenClaw
Not every automation task requires computer use. Use it when:
- No API or programmatic interface exists
- An API exists but is poorly documented or unstable
- You're integrating with legacy systems that predate modern APIs
- The workflow is genuinely interactive and requires visual feedback
Prefer API-driven automation when possible: it's faster, more reliable, and easier to audit. Computer use should be a fallback for situations where APIs aren't practical.
Competitive Landscape
RPA vendors (UiPath, Automation Anywhere, Blue Prism) have dominated desktop automation, but they require extensive configuration and maintenance. OpenClaw agents with improved computer use, powered by Vercept's expertise, become genuinely competitive alternatives: more flexible, cheaper to operate, and capable of handling variation without reconfiguration.
The Vercept acquisition is Anthropic's signal that they're serious about going beyond text-based AI. Computer use automation is a multi-billion-dollar market, and Anthropic is positioning Claude to compete directly with established RPA platforms.
Implementation Recommendations
For OpenClaw users considering computer use automation:
- Start with small, low-risk pilots: basic web form filling, simple data extraction
- Implement strong monitoring and logging to detect failures or unexpected behavior
- Maintain fallback procedures: if agent automation fails, what happens? Who gets notified?
- Build gradually: as reliability improves, expand to more complex workflows
- Test extensively in isolated environments before touching production systems
- Train teams on computer use security implications—it's more powerful and requires more careful governance than text-based tools
The Vercept acquisition signals that OpenClaw agents are evolving from assistants and advisors to genuine automation platforms. This is exciting for efficiency and capability, but it also demands that we think carefully about governance, security, and human oversight. The best OpenClaw deployments will leverage computer use strategically while maintaining clear boundaries and approval workflows.