The Plan Review MCP checks your AI agent's plan against your codebase's conventions — before a line is written. PR review scores the diff before it merges. Nine architectural detectors. One score. Learn more →
CI proves the code runs — it can't prove the code belongs. Bloated, AI-authored code multiplies token costs ~2× every 60 days. The cheapest place to stop a duplicate service or an unsafe migration is before it's written — not in a line-by-line PR review after the fact.
Code calling modules that don't exist in your repo.
The third competing payment workflow, under a new name.
The agent didn't know a helper already existed.
Each PR follows its own convention, not yours.
Five layers of wrapper around one call site.
Every drift above is invisible in a diff, invisible to linters, and compounds across hundreds of AI-authored PRs. A line-by-line review can't catch what a reviewer can't see.
44% of teams already run AI review on PRs. None of them review the plan — the one place drift can still be stopped for free.
CodeRabbit & Copilot own inline PR comments. Greptile owns codebase context. Sonar & Aikido own quality and security gates. Nobody else owns the plan, the entropy score, or the ADRs.
| Capability | VibeCop | CodeRabbit | Greptile | Qodo | Copilot | Sonar | Aikido |
|---|---|---|---|---|---|---|---|
| Pre-code plan review (MCP) | ✓ | ✗ | ✗ | ◐ | ✗ | ✗ | ✗ |
| Architectural entropy score (AII) | ✓ | ◐ | ◐ | ◐ | ✗ | ◐ | ✗ |
| ADR-aware review | ✓ | ✗ | ◐ | ◐ | ✗ | ✗ | ✗ |
| Knowledge graph + graph Q&A | ✓ | ✓ | ✓ | ◐ | ◐ | ✗ | ✗ |
| One rolled-up ticket per PR | ✓ | ◐ | ◐ | ◐ | ◐ | ◐ | ✓ |
| Inline PR comments | ✗ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Learns from reviewer feedback | ✗ | ✓ | ✓ | ◐ | ◐ | ✗ | ✗ |
| Deep security (SAST/SCA) | ◐ | ◐ | ◐ | ◐ | ◐ | ✓ | ✓ |
We skip inline comments & feedback-learning by design — we sell the review nobody else runs, not a louder version of theirs.
That "review the plan" column is a remote MCP server. Point Claude Code, Codex, Cursor, or any MCP-capable agent at it and every plan gets checked against your codebase before a line is written.
Claude Code, Codex, Cursor, or any MCP-capable agent writes its implementation plan from your PRD or spec.
The agent calls vibecop_check — the plan is reviewed against YOUR codebase's conventions, ADRs, and existing utilities.
Would-be duplication, hallucinated abstractions, and pattern drift are caught while they're still sentences, not diffs.
The codebase stays lean, so every future agent run loads less context, drifts less, and costs less.
A conservative one-year snapshot: ~4.8B wasted tokens ≈ ~$19K — and because drift compounds (bloat roughly doubles every ~60 days), that tax accelerates. Prevention keeps it near zero.
Illustrative example only. Your numbers depend on team size, model mix, and how much drift has already accumulated.
claude mcp add --transport http vibecop https://api.gradus.maiife.ai/mcp/vibecop --header "Authorization: Bearer <TOKEN>"Every plan can mint an MCP token in Settings → Plan Review MCP.
FULL SETUP GUIDE →A single score from 0–100 that captures your codebase's structural health. Four weighted axes: Consistency, Reuse, Simplicity, and Structure.
An LLM synthesis layer that reads all detector signals, applies your team's ADRs, and produces a narrative recommendation — approve, request changes, or reject — with cited evidence.
Full dependency visualization with fan-in/out analysis, coupling metrics, and cycle detection. See how a single PR impacts the broader architecture before it merges.
Monitor drift velocity over time. Spot when duplication creeps in, when abstractions diverge, and when pattern variance accelerates — before it becomes tech debt.
Link your GitHub repository. VibeCop installs as a lightweight integration — no CI changes, no config files, no pipeline modifications.
VibeCop automatically triggers on every pull request. Nine architectural detector agents analyze the diff against your full codebase graph in under 90 seconds.
The Judge synthesizes all signals into a single narrative — with an Architecture Integrity score, cited findings, and a clear approve/reject recommendation.
Detects code structures that exist without justification — interfaces nobody implements, factories that produce one product.
Finds logic that does the same thing in different words. Not copy-paste — conceptual overlap that fragments your domain model.
Identifies unnecessary abstraction layers, premature generalization, and design patterns applied where a function would suffice.
Spots when new code deviates from your team's established conventions — error handling style, naming, file structure.
Catches hand-rolled implementations of functionality that already exists in your codebase or standard library.
VibeCop is Maiife's Architecture Integrity & Product Intelligence platform. It scores repositories with a 0–100 Architecture Integrity Index and flags architectural entropy in AI-generated code — hallucinated abstractions, duplicated logic, and pattern drift — before it merges.
The Architecture Integrity Index is a 0–100 score of your codebase's structural health, computed as 100 minus entropy impact. It weighs four axes: Consistency, Reuse, Simplicity, and Structure.
VibeCop runs nine architectural detector agents across four axes — hallucinated abstractions, semantic duplication, over-engineering, pattern drift, reinvented utilities, architectural complexity, layer violations, orphaned modules, and reintroduced bugs — plus seven static hygiene layers (dependency CVEs, dependency health, secrets, SAST, infrastructure-as-code, vendored drift, magic numbers) tracked separately from the index.
Yes. VibeCop connects directly to GitHub repositories and reviews pull requests. Webhook auto-reviews trigger a full analysis on every PR open and sync.
VibeCop is free for everyone until December 31, 2026 — no card required, and nothing is billing. Pricing has not been finalised and will be announced before any plan starts billing.
Linters check line-level style and syntax. VibeCop analyzes architecture-level entropy — whether new code drifts from your team's established patterns, duplicates existing logic semantically, or adds unjustified abstraction — using a full code graph of your repository.
A remote MCP server that reviews your AI agent's implementation plan against your codebase's conventions before any code is written. It works with Claude Code, Cursor, Codex, and any MCP-capable agent over streamable HTTP with a Bearer token minted in VibeCop Settings.
Mint an MCP token in VibeCop Settings, then add the server: claude mcp add --transport http vibecop https://api.gradus.maiife.ai/mcp/vibecop --header "Authorization: Bearer <TOKEN>" — or add the equivalent mcpServers entry to your agent's MCP config. Setup guide at vibecop.maiife.ai/mcp.
Free trial includes AI credits, full detector suite, and architecture graph.