We’ve been working on significant updates to the core engine of Mechasm.ai. Our focus this cycle has been on solving access challenges and making debugging clearer and faster.
1. Isolated VPN Tunnels
Testing public URLs is straightforward, but securely testing internal staging environments from the cloud has always been complicated. You effectively have to choose between glaring security vulnerabilities (allow-listing broad IP ranges) or relying on complex, often flaky tunnel infrastructure.
The Problem: Access vs. Security
To test an internal staging application, cloud agents traditionally need broad access to your network. This creates a security risk. Furthermore, if you are running multiple projects, managing overlapping IP ranges and permissions becomes a nightmare.
The Solution: Per-Project Network Encapsulation
To solve this, per-project network encapsulation is now live. We have re-architected how our test runners handle network traffic:
- Ephemeral Namespaces: When a test runs, the system creates an isolated network environment (using Linux namespaces) exclusively for that project.
- The "Bubble" Architecture: The VPN connection happens entirely within this secure bubble. The browser traffic for that specific test is routed through this encrypted tunnel.
- Full Protocol Support: We now support both OpenVPN and WireGuard protocols out of the box.
This allows a test in Project A to tunnel into a specific internal VPC, while a test in Project B remains completely isolated on a separate network. It ensures there is zero leakage between projects and keeps your internal infrastructure secure without complex firewall management.
2. Smarter Debug Traces
When an automated test fails, the first question is always "Why?" Standard automation traces can be heavy, noisy, and difficult to parse. Finding the exact moment a logic error occurred often means sifting through gigabytes of raw browser events and redundant screenshots.
The Problem: Signal to Noise
Traditional traces dump everything—every network ping, every pixel shift, every memory buffer. This results in heavy files that are slow to load and hard to read. You spend more time fighting the tool than fixing the bug.
The Solution: Action-Aware Tracing
We re-engineered our trace system to cut through the noise and focus on what actually matters:
- Optimized Capture: The engine now actively optimizes trace files to remove partially loaded or redundant frames.
- Intelligent Snapshots: Instead of random intervals, we inject high-quality snapshots exactly when a logical step completes (e.g., "After clicking 'Submit'", "After filling the form").
- Instant Replay Viewer: We introduced a new, client-side trace viewer that parses everything directly in your browser. This provides a timeline that synchronizes the Screen State, Console Logs, and Agent Actions, allowing you to replay exactly what happened immediately without waiting for server processing.
3. Other Improvements
Alongside these major features, we've deployed several quality-of-life improvements:
Intelligent Validation
Better feedback starts before the test even runs. The agent now performs pre-flight validation on your test steps. It intelligently distinguishes between valid browser actions and unsafe system operations (like executing shell commands or accessing the file system). It guides users with constructive feedback to ensure tests are automatable and safe before execution begins.
Improved Accuracy
We’ve fine-tuned the underlying model to better identify UI elements, drastically reducing flaky interactions on dynamic applications.
Stability & Performance
Several backend enhancements have been deployed to improve overall platform uptime and significantly speed up test execution times.
What's Next?
These updates are part of our commitment to building an agent that understands your infrastructure just as well as it understands your code.
These features are live now. Log in to Mechasm.ai to try them out.