v1.6.0 — Multi-LLM consensus code review

Catch stronger signals.
Ignore weak ones.

8 AI models. One trusted output.

Run GPT-4o, Claude, Gemini, Groq, Cerebras, Mistral, OpenRouter, and Ollama against your code in parallel. Only findings that hold up across providers reach your report — with explicit confidence scores and disagreement detection.

Built as a second validation layer before release or client handoff.

Get Pro — €19 Try Free

Windows 10/11 64-bit·Portable .exe·BYOK·No install

nexaverify_report.json

// scan complete · 3 providers · balanced mode · v1.6.0_DEV

{

  "health_score": 40,

  "mode": "balanced",

  "providers": ["gemini", "groq", "cerebras"],

  "total_issues": 99,

  "critical": 0, "high": 0, "medium": 98, "low": 1,

  "reliability": "partial_confidence"

}

Who it's for

When a single AI review isn't enough

If you ship code that AI helped write, single-model review hides risk. Multi-LLM consensus surfaces it.

AI-heavy codebases

When Cursor, Copilot, or ChatGPT writes large parts of your code, review speed goes up — and so does uncertainty. NexaVerify adds a cross-check before that uncertainty reaches production.

Pre-delivery audits

Add a validation pass when stakes are high — client handoff, production release, or sensitive code paths. Review confidence scores before shipping.

Freelancers & consultants

Justify audit fees with a professional multi-LLM consensus report. Concrete deliverable instead of vague "I reviewed it" assurance.

Best used before client delivery

Best used before production release

Best used after heavy AI-generated code

Best used on auth, shell, or runtime-sensitive logic

Why not just one LLM

Single-model trust is broken in 2026

96% of developers don't fully trust AI-generated code. Only 48% verify systematically. NexaVerify is for the other 52% who know they should — but don't have time.

Single-model review

One LLM returns findings. Some are real. Some are hallucinations. You can't tell which is which — so you either trust everything (waste time on false positives) or ignore everything (miss real bugs).

Multi-provider consensus

Multiple LLMs analyze the same code independently. NexaVerify compares outputs — issues confirmed across providers get higher confidence. Isolated findings get lower. Disagreements are surfaced, not hidden.

Scan modes

Three depths, one workflow

Quick

Fastest sanity pass. Small checks, quick signal.

Balanced ★

Best default. Speed, signal quality, and cost balanced.

Deep

Pre-delivery audits, client work, sensitive code.

How it works

From folder to validated report

STEP 01

Map

Select project. NexaVerify builds a local view: files, chunks, stack, hotspots.

STEP 02

Validate

Multiple providers analyze in parallel. Consensus engine filters weak or isolated signals.

STEP 03

JSON Repair

Truncated or malformed LLM responses are automatically repaired — bracket matching, fallback extraction, and structural validation before parsing.

STEP 04

Deliver

HTML for humans. JSON for automation, archiving, run-to-run diffing.

JSON Repair schema — automatic recovery of truncated LLM responses

JSON Repair pipeline — automatic bracket matching, extraction, and validation of provider responses.

After a scan

What you actually receive

Validated issue list

Findings ranked by severity, weighted by how many providers agreed.

Confidence per finding

Single-provider: lower confidence. Multi-provider agreement: higher confidence.

Disagreement flags

When providers disagree on severity, it's surfaced instead of hidden.

Dual output

HTML report for human review. JSON export for automation and diffing.

Provider health

See which providers succeeded, which failed, and why — in every report.

Prompt presets

Targeted review angles: bugs, security, performance, QA.

BYOK setup

Running in 30 seconds with your own keys

Bring your own API keys. No NEXADiag account. No proxy. Keys stay on your machine.

01Get a free key

Gemini's free tier signs up in 30 seconds at aistudio.google.com/apikey — no phone needed. Or use GPT, Claude, Groq — any provider you already have.

02Paste in NexaVerify

Open the app, paste your key in Settings. Done. Key stays on your machine, never sent to NEXADiag servers.

03Scan a folder

Browse to your project, click Analyze. HTML report opens in your browser. That's it.

Real output

Not mockups — actual scan results

A real scan of the NexaVerify v1.6.0 codebase itself — 3 free-tier providers (Gemini, Groq, Cerebras) in parallel. 99 issues found, 0 critical.

Health: 40/100 99 issues surfaced 98 medium · 1 low 3 free LLMs in parallel Partial confidence

Real scan · Balanced mode · 3 providers (Gemini, Groq, Cerebras) · Project: NexaVerify v1.6.0_DEV itself.

→ Open the actual live report ▼ Show full breakdown

Health

Critical

High

Medium

Low

Total

Real issues found in the NexaVerify source code — click "Open the actual live report" for all 99.

File	Type	Severity	Detected by	Confidence
main.py:118	bug	MEDIUM	gemini	36%
_init_settings() not wrapped in try-except — crash before logging is initialized
constants.py	bug	MEDIUM	cerebras	36%
WINDOW_TITLE references undefined variables — NameError on import
analysis_support_v160.py:124	bug	MEDIUM	cerebras	36%
Missing return value — callers receive None instead of AnalysisPipelineResult
arbiter.py	bug	MEDIUM	cerebras	36%
build_consensus_output doesn't return ConsensusOutput — callers receive None
consensus.py:165	bug	MEDIUM	cerebras	36%
Wrong argument passed to _provider_weight — potential TypeError

Real bugs found in NexaVerify's own code. The tool scanned its 26,000+ line codebase and surfaced these real issues — missing error handling, undefined variables, missing return values, and type mismatches.
Confidence: Partial (3/4 provider slots filled — free tier limit). Gemini, Groq, Cerebras all returned data.

This is what you get: real issues, real confidence scores, real provider tracking. Not a demo. Not a simulation.

Get Pro — €19

Provider activity

Visibility on every provider in every scan

From the same real v1.6.0 scan: each provider reports independently, the engine tracks who returned what, and any failure is flagged instead of silently dropped.

gemini18 results

groq91% success

cerebras97% success

↓

Consensus engine — 3 of 4 provider slots filled (free tier)

Partial confidence flagged automatically. All 3 free providers returned usable data.
The 4th slot was unused due to free tier limit (3 providers max) — clearly indicated in the report.

Interface

Desktop app — empty state before scan

All indicators at 0 until a scan is run.

NexaVerify multi-LLM code review running on a project folder

Live view: providers analyzing in parallel.

Honest limits

What NexaVerify is — and what it isn't

✓

Local execution. Windows 10/11 64-bit. No server between you and the LLM providers.

✓

Your API keys. Stored locally. Sent directly to provider endpoints. No NEXADiag proxy.

✓

Code not stored. Results generated in session. No code persists on disk.

⚠

Internet required for analysis. Local means no NEXADiag server. LLM API calls need connectivity.

⚠

Not a replacement for human review. A validation layer — especially valuable where providers disagree.

Active development

A tool that grows with its users

✓ Available today

Multi-provider consensus with 8 AI engines. Quick / Balanced / Deep modes. HTML + JSON + SARIF 2.1.0 reports. Disagreement detection. Local-first with your API keys. Ollama offline support. Proof of Determinism.

⚙ Currently improving

Smarter consensus weighting. Better grouping of related issues. Improved performance on large codebases. Scan history and run-to-run comparison.

→ Built with users

Early users directly influence priorities. Not a closed SaaS — a tool that grows with real workflows. Buyers today shape what ships next.

No hype roadmap. Only what is actively being built.

Curious how it actually works before buying?

→ Read the full user manual

Pricing

One payment. Lifetime access.

No subscription. No lock-in. No "starter plan" upsell traps.

€19one-time

Pay once. Use forever. No subscription.

★ Early-bird price for first buyers. Will increase after launch phase.

Lifetime license. Machine-bound. Free updates as the tool evolves.

Get Pro — €19 Try Free

✓ Multi-provider consensus engine

✓ Quick / Balanced / Deep scan modes

✓ HTML + JSON reports

✓ Confidence scoring + disagreement detection

✓ Prompt presets (bugs, security, performance, QA)

✓ Diagnostics export from Control Center

Free tier: limited analyses/day · resets at local midnight · fewer files per scan · reduced provider slots.

Feature	Free	Pro (€19)
AI providers	3 (Gemini + Groq + Cerebras)	All 8
Analyses per day	3	Unlimited
Files per scan	10	200
Consensus Engine	✓	✓
HTML + JSON reports	✓	✓
SARIF 2.1.0 export	✓	✓
Ollama local support	✓	✓
Proof of Determinism	✓	✓

Technical writing

Articles from building NexaVerify

Real bugs, real fixes, no hype. Published while building the tool.

Why JSON.parse() Fails Silently on Truncated LLM Responses (And What I Did About It) dev.to — bracket repair, isinstance guards, why production LLM code needs more paranoia than normal API code.
Stop Copy-Pasting AI Code — The 6-Step Validation Checklist for Devs dev.to — code review, edge cases, dependency validation, security, and cross-LLM checking in practice.

FAQ

Frequently asked questions about multi-LLM code review

Why multiple providers instead of one? ▼

Single-model reviews hallucinate and miss things. With multiple providers, NexaVerify compares signals — weak false positives get lower confidence. Real issues confirmed across providers get higher confidence.

When should I use Deep vs Balanced? ▼

Balanced is the daily default. Use Deep before delivery, on sensitive code, or when a client is paying for a thorough audit.

What if a provider fails mid-scan? ▼

The scan continues with the providers that responded. The report is generated with a "Partial confidence" label so you know the consensus was based on fewer signals. Keep at least 2 active providers to absorb individual failures.

What is the JSON output for? ▼

Archiving, automation, historical tracking, or comparing multiple runs. HTML for human reading.

Does it store my code or API keys? ▼

No. API keys stay local. Code is sent directly to provider endpoints. No telemetry, no tracking.

What if I change PC? ▼

License is machine-bound. Request a reset via support with proof of purchase.

Linux or Mac support? ▼

Windows 10/11 64-bit native. Wine on Linux (experimental). macOS not yet available.

Will the tool keep evolving after I buy? ▼

Yes. NexaVerify is in active development. Buyers receive updates as the consensus engine, reporting, and provider handling improve. Your feedback directly shapes the roadmap.

Catch stronger signals.Ignore weak ones.