I'm building Kenwea Code, an independent verification layer for AI-assisted software development.
I'd love to hear from developers who use AI to write code every day, including vibe coders.
Here's a problem I keep running into: one model writes the plan, writes the code, and then reviews its own work. It can miss the same mistake twice. And by the end, it's often difficult to know what was spent, which decisions were made, and why.
Kenwea takes a different approach: models from different providers check each other's work.
For example, Claude might write a specification while GPT and Gemini independently review it.
But simply asking multiple models for opinions isn't enough. Kenwea uses a few rules to make those reviews meaningful:
Blind first round: Each model works independently, without seeing the others' answers.
Evidence-based objections: Every objection must quote the exact text it challenges and explain the potential consequence. Vague criticism is filtered out.
Independent arbitration: If reviewers disagree, another model acts as an arbiter. If there's still no resolution, you see both positions and make the final call.
Cost and decision tracking: Every run starts with a cost estimate and spending limit. Model calls, objections, and decisions are recorded for later verification.
Kenwea doesn't aim to replace Claude Code, Codex, or Gemini CLI.
They remain the coding agents. Kenwea sits above them as an independent planning, review, and accountability layer.
The goal is simple: the model that produces the work shouldn't be the only one deciding whether that work is good.
Where things stand
I'm currently building the first component:
harness plan "your idea"
The initial workflow will:
Take your idea through a multi-model panel.
Ask up to 20 clarification questions.
Generate project documents: requirements, architecture, threat model, test plan, and task list.
Have models from different providers cross-review those documents.
Deliver the approved documents together with cost information and an auditable decision log.
Future releases will add task orchestration, cross-provider code review, an MCP server, GitHub Actions integration, browser-based animation testing, and eventually a desktop control room.
I'd genuinely like your feedback
When an AI coding tool gives you a plan or a diff, what do you still check manually because you don't trust it?
Would you pay for a second opinion from an independent model on every change, or only on high-risk changes?
Where would you actually use this: CLI, MCP inside your editor, or GitHub Actions on pull requests?
What's one feature you've wished your AI coding tool had but haven't found yet?
Blunt feedback is welcome. Even "I wouldn't use this, and here's why" would be valuable.
I'm building this in public and would rather discover what developers actually need before building the wrong thing.
Top comments (0)