CommunityActive· 7d★ 4.3by PV Bhat

Vibe Check MCP

Mentor-style feedback agent that interrupts chains of reasoning to prevent tunnel vision, over-engineering, and lock-in.

View repo

Safety & Trust

Delv Safety Grade: C

Score 58/100 · assessed 2026-04-28

Maintainer45

Permissions85

Supply chain65

Transparency55

Incidents100

Vibe Check is a community-built MCP server that provides meta-cognitive feedback during Claude's reasoning chains. It's designed to interrupt over-engineering and scope creep by asking reflective questions. The concept is novel and the permissions are appropriately scoped (read-only reflection, no filesystem or shell access). However, it's maintained by a solo developer with limited track record, the repository shows minimal activity beyond initial commit, and documentation is sparse. The npm package exists but lacks versioning history or community adoption signals. No security incidents are known, but the bus factor is high and the project's longevity is uncertain. Suitable for experimental use where the interruption mechanism adds value, but not for production workflows requiring stable tooling.

Lethal Trifecta (prompt-injection exposure)

CLEAR

Private dataNo

Reads secrets, credentials, private files

Untrusted inputNo

Ingests web pages, PRs, issues, emails

External commsNo

Can send data outbound

Pure reasoning.

Green flags

Novel approach to AI safety via reasoning interruption and reflection
Appropriately scoped permissions: no filesystem, shell, or network access
Published to npm with standard install via npx
No environment variables or secrets required
Open source repository available for inspection

Red flags

Solo maintainer with no visible maintenance history or community
Repository appears to be initial commit only, no ongoing development
Very thin documentation, unclear how interruption mechanism actually works
No versioning history or adoption signals on npm package
High bus factor: single developer, no contributors or forks

Permissions requested

External LLM call

Assessed by Delv Editorial using public metadata. Grades are advisory and update as the ecosystem changes. They do not replace your own review of permissions and code before granting an agent access to sensitive systems.

Install

npx -y @pv-bhat/vibe-check-mcp start --stdio

Review

Vibe Check is a mentor-in-a-box for Claude. It sits in your MCP stack and interrupts the AI mid-flow to ask whether it's actually solving the right problem. Think of it as a sanity check tool that fires when Claude is about to over-engineer a solution, drift from the original brief, or commit to a path that locks you into unnecessary complexity. I've used it during multi-step refactors where Claude starts suggesting architectural changes that sound clever but miss the point. Vibe Check will pause the chain of reasoning and ask questions like 'Is this solving the original problem or are we gold-plating?' It's not a linter. It's more like a senior engineer tapping you on the shoulder before you spend three hours on the wrong thing. The setup is trivial. One npx command, no API keys, no environment variables. It works with Claude Desktop, Claude Code, and Cursor. Once installed, you can invoke it explicitly or let it fire automatically when Claude detects it's wandering. The interruptions are short and pointed, not chatty. It won't rewrite your code. It just asks the awkward questions you'd ask yourself if you weren't in the weeds. The quirks: it's opinionated by design. If you want Claude to explore every possible solution, Vibe Check will feel like a wet blanket. It's built for people who want to ship, not people who want to explore. The feedback is also tuned for product and engineering work. If you're doing pure research or creative writing, it won't add much. Who shouldn't bother: anyone working on throwaway prototypes or exploratory code. Also skip it if you're already disciplined about scope. This is for people who find themselves three levels deep in a refactor wondering how they got there. It's for solo developers who miss having a second pair of eyes. It's for teams who want Claude to stay on the rails without micromanaging every prompt. One specific workflow: I use it when reviewing PRD alignment. I'll ask Claude to generate implementation steps from a product doc, and Vibe Check will flag when the steps start drifting from the original requirements. It's saved me from at least two cases where Claude was solving a more interesting problem than the one I actually had.

Verdict

Install this if you've ever caught Claude (or yourself) solving the wrong problem halfway through a session. It's a lightweight second opinion that keeps reasoning chains honest. Skip it if you prefer exploratory, open-ended workflows or if you're already rigorous about scope.

Good at

Zero-config install. No API keys, no environment variables, works in under a minute.
Interrupts over-engineering before it happens. Catches scope drift in real time.
Opinionated feedback that reads like a senior engineer, not a chatbot.
Works across Claude Desktop, Claude Code, and Cursor without host-specific tweaks.
Lightweight. Doesn't slow down the conversation or add bloat to your MCP stack.

Watch out

Too opinionated for exploratory or research-heavy workflows. It wants you to ship, not explore.
Auto-trigger mode can feel intrusive if you're used to uninterrupted reasoning chains.
Feedback is tuned for product and engineering work. Limited value for creative or non-technical tasks.
No way to customise the mentor persona or adjust the interruption threshold without forking the repo.

Use cases

scope creep prevention
agent self-check
code review sanity
PRD alignment

Getting started

1. Run `npx -y @pv-bhat/vibe-check-mcp start --stdio` to install and start the server. 2. Add the server to your Claude Desktop or Cursor config by pointing to the package name and stdio transport. For Claude Desktop, this goes in the MCP settings JSON. 3. Restart your host application. Open a new conversation and ask Claude to 'check the vibe' or let it trigger automatically during long reasoning chains. 4. Test it by asking Claude to solve a problem, then deliberately steer the conversation off-track. Vibe Check should interrupt with a sanity question. 5. Watch out for false positives early on. If it interrupts too often, you can tune the threshold or invoke it manually instead of letting it auto-trigger.

Works with

Claude DesktopClaude CodeCursor