Architecture Review: cubic 2.0

cubic 2.0 claims to be “Instant, context-aware PR reviews right in GitHub.” With the rise of “AI Engineers” (like Devin or Cursor), the bottleneck has shifted from writing code to reviewing it. Cubic aims to solve this by acting as an automated senior engineer that reviews your PRs before a human does.

🛠️ The Tech Stack

Cubic 2.0 goes beyond a simple “diff-to-LLM” script. It employs a sophisticated orchestration layer designed to understand the intent and context of a codebase, not just the lines changed.

• Core Integration: It operates primarily as a GitHub App. This allows it to listen to pull_request events via webhooks, fetch diffs, and post inline comments directly using the GitHub API.
• LLM Orchestration (The “Brain”): Unlike single-model wrappers, cubic utilizes a Model Router. It dynamically selects the best model for the task-likely routing complex reasoning to Claude 3.5 Sonnet (or “Claude Code”) and faster, simpler linting tasks to lighter models (like GPT-4o mini).
• Context Engine (RAG + Graph): To achieve “context awareness,” cubic likely indexes the repository into a vector database or dependency graph. This allows the AI to understand that a change in utils.ts might break a function in auth.ts, even if auth.ts wasn’t touched in the PR.
• Agentic Sandbox: A standout feature is “Background Agents.” Cubic spins up ephemeral, secure sandboxes to actually run or verify fixes (using tools like Claude Code) rather than just hallucinating a suggestion. This is a critical step up from standard text-generation tools.
• Security: SOC 2 Type I compliance and ephemeral processing (data is wiped after review) suggest a mature architecture designed for enterprise security requirements.

💰 Pricing Model

Cubic follows a standard SaaS developer tool pricing strategy with a generous entry point.

• Free Tier (Freemium):
  • Public Repos: Completely free (unlimited reviews). This is a smart growth hack to capture the Open Source market.
  • Private Repos: Includes 20 AI reviews per month for free. This is sufficient for hobbyists or solo devs but will quickly cap out for active teams.
• Paid Tier ($30/seat/month):
  • Unlimited AI reviews.
  • Custom rules (e.g., “Always use zod for validation”).
  • Stacked PR support (critical for high-velocity teams).
  • Priority support and “Background Agents” capabilities.

⚖️ Architect’s Verdict

Is cubic 2.0 just a wrapper? No.

While v1.0 iterations of such tools were often simple wrappers around OpenAI’s API, cubic 2.0 represents “Deep Tech” in the DevTool space. The differentiation lies in the Context Engine and Sandboxed Agents.

A wrapper simply asks an LLM, “What is wrong with this code?” Cubic, however, appears to be building a semantic map of your codebase and validating its own suggestions via agents. The ability to “learn” from previous PR comments and enforce team-specific patterns moves it from a “spell checker” to a “process automator.”

Developer Use Case: Perfect for teams drowning in PRs. If your senior engineers spend 50% of their time nitpicking variable names or catching missing error handlers, cubic is a no-brainer. It acts as the “Level 1” support, ensuring that by the time a human looks at the PR, the trivial issues are resolved, and the context is clear.

Rating: Production Ready. The SOC 2 compliance and “Background Agent” architecture make it a viable enterprise tool, not just a toy.