Sentinel DV: An AI-Powered Intelligent Analysis Gateway for Chip Verification

Sentinel DV is an open-source MCP server designed specifically for the chip verification domain. It aims to enable large models to safely access SystemVerilog/UVM verification data for fault localization and coverage analysis, while avoiding exposure of sensitive logs or simulator control. Its core value lies in addressing sensitive information leakage and security risks when using AI for verification assistance.

Chip design verification is one of the most time-consuming processes in the semiconductor industry. Modern CPUs contain billions of transistors, requiring verification engineers to write massive test cases and run thousands of regression tests. In traditional workflows, engineers have to switch between multiple tools, limiting efficiency. The idea of AI-assisted verification emerged as a solution, but verification data often contains sensitive information (e.g., IP details, keys), so directly allowing AI to access raw logs or simulators poses security risks.

Sentinel DV is based on the Model Context Protocol (MCP). Its core design philosophy includes:

1. Read-only access: Prohibits write operations or simulator control to ensure AI errors or attacks do not damage the verification environment or leak data;
2. Structured data: Converts UVM logs, coverage reports, etc., into a unified typed schema, preserving key information while denoising and desensitizing;
3. Security-first architecture: Built-in mechanisms for automatic desensitization, path sandboxing, response size limits, and source tracing.

Sentinel DV uses a layered architecture:

1. Schema layer: Defines typed schemas (e.g., TestCase, FailureEvent) with version management to ensure interface compatibility;
2. Adapter layer: Supports UVM, cocotb, native SystemVerilog, and mainstream simulators (VCS, Xcelium, etc.);
3. Index and storage layer: Uses DuckDB as the backend for efficient querying and storage of large-scale verification data;
4. Tool layer: Exposes MCP interfaces (e.g., runs.list, failures.list) for AI Agents to call.

Sentinel DV's application scenarios include:

1. Intelligent fault localization: AI queries failure event classifications and evidence via interfaces to quickly narrow down root cause ranges;
2. Coverage analysis: Answers questions about insufficient coverage points, untriggered assertions, etc.;
3. Regression comparison: Compares differences between two regressions to identify new failures or coverage change trends;
4. UVM topology understanding: Extracts the hierarchical structure of the verification environment to help new team members get up to speed quickly.

Sentinel DV's security measures include:

1. Automatic desensitization engine: Identifies and replaces sensitive information (e.g., AWS keys, GitHub tokens, emails, etc.);
2. Deterministic output: Normalized processing ensures the same input produces the same output;
3. Evidence reference mechanism: Key facts are accompanied by the location of original verification artifacts for easy auditing;
4. Sandbox and boundaries: Controls the directory range accessible by AI via configuration to prevent path traversal attacks.

Sentinel DV represents an important direction for AI-assisted chip verification: enabling AI to be an intelligent assistant for engineers, providing data insights within secure boundaries. For verification teams: It lowers the threshold for AI adoption without requiring modifications to existing workflows. For AI developers: It provides standardized interfaces to access the verification domain. In the future, it is expected to free engineers from tedious data organization tasks, allowing them to focus on creative work and address verification bottlenecks caused by the growing complexity of chips.

Quickly experience Sentinel DV:

1. Installation: In a Python 3.10+ environment, execute pip install sentinel-dv;
2. Configuration: Edit config.yaml to specify the verification artifact directory and enabled adapters;
3. Startup: Run python -m sentinel_dv.server --config config.yaml;
4. License: Open-source under the Apache 2.0 license. Contributions and feedback are welcome.