Hey devs ๐ โ especially folks in Singapore juggling multiple AI vendors (DeepSeek for cost, Claude for reasoning, GPT for everything else).
If you've done the manual copy-paste dance โ draft in one model, paste into another asking "check this for me", paste back, repeat โ bayze-mcp turns that dance into one sentence in your AI client (Cherry Studio / Cursor / Claude Desktop โ anything MCP):
Use review_then_revise for: [your important draft]
Model A drafts โ Model B (different vendor) reviews with structured JSON verdict โ Model C revises only if needed. Every step logs provider, model, tokens, cost. Or just say it in plain language โ v1.4.1 added natural-language routing, so "cross-check this for me" makes the client pick the right workflow.
Five collaboration patterns
- Cross-vendor review (~$0.001/run): A drafts โ B reviews โ C revises. Caught a missing TLS handshake detail that single-model self-review missed.
- Parallel + judge (~$0.0013/run): N models answer, judge picks by hard constraints.
- MoA aggregation (~$0.006/run): synthesizer merges the best of N answers.
- Revision loop with escalation: review โ revise โ re-review โ escalate to a stronger model.
- Adversarial debate (~$0.02/run): two debaters, arbiter verdict that named and excluded fabricated "audit data" both sides invented.
What 6 rounds of real testing taught us
- No model can count its own words. Off by 3โ38 chars/words across 5 samples. So the engine counts deterministically โ never trusts model self-reporting.
- Rewrite-type steps systematically lengthen text (+59%, +74%, 1.66ร). If you need hard length caps, use selection, not synthesis.
- Both debaters fabricated "audited invoices" and benchmark numbers. The arbiter now names and excludes unverifiable claims.
- The engine saved the day repeatedly: client timeout reported while engine finished fine (local journal for reconciliation); verdicts exceeding length caps get self-calibrated compress-retry instead of silent truncation.
For teams in Singapore specifically
- Local-first: keys stay in your env vars, no telemetry, no cloud dependency. Works fully offline with local models (Ollama etc. via OpenAI-compatible endpoint).
- All major vendors work: OpenAI, Claude, Gemini + China vendors (DeepSeek/Qwen/Doubao) โ useful for cost optimization across vendors.
- MIT license. 220 tests. TypeScript strict. Node โฅ22.
Quick start
npm install -g bayze-mcp
bayze init # interactive wizard: pick providers, paste keys, generates config + client registration
Then in Cherry Studio / Cursor: add MCP server โ done. Or Claude Desktop: bayze init auto-registers into your existing config.
Links:
- GitHub: https://github.com/caijianxiong2007-prog/bayze-mcp
- npm: https://www.npmjs.com/package/bayze-mcp
- Full debate debugging log (previous post): https://dev.to/caijianxiong2007prog/we-made-two-ais-debate-and-hired-a-third-as-the-arbiter-6-test-rounds-7-bugs-fixed-4c5
Happy to answer questions โ especially if you've caught your models fabricating data too. We keep a list.
Top comments (1)
Cross-vendor review helps, but different vendors still miss the same thing when the error sits in the source material they were all given. I'd log, for each verdict, which claims the reviewer actually checked against the source. Otherwise an approval can just mean nothing looked obviously wrong.