nvhive 0.31.1

NVHive — Multi-LLM orchestration platform with intelligent routing, hive consensus, and auto-agent generation

nvHive

One command. Every AI model you have. Automatically assembled into the best team for each task.

nvh "What is a binary search tree?"              # → answers (single best advisor)
nvh "Fix the timeout bug in council.py"          # → auto-detects coding task → agent mode
nvh "Should we use Redis or Postgres?"           # → auto-detects debate → council (3+ advisors)
nvh "take a screenshot and describe my desktop"  # → desktop agent (vision + tools)
nvh "setup comfyui"                              # → agent installs, configures, launches

---

Get Started

pip install nvhive
nvh                    # first-run setup auto-detects GPU, installs local AI, configures providers
nvh "your question"    # just ask — nvhive figures out the rest

# Optional extras
pip install "nvhive[vision]"      # desktop agent: screenshot, click, type, scroll
pip install "nvhive[browser]"     # headless browser automation (playwright)
pip install "nvhive[all]"         # everything

On first run, nvh launches a guided 3-step setup — GPU detection, provider keys, local model pulls. Works immediately with local models (no signup needed). Every step is skippable. Run nvh setup anytime to reconfigure.

GPU tier → model recommendations:

VRAM	Text Model	Vision Model	Behavior
0 GB (no GPU)	Cloud only	Cloud fallback	Free tiers first (Groq, LLM7, GitHub)
4-8 GB	nemotron-mini	moondream	Basic local + desktop agent
12-16 GB	qwen2.5-coder:7b	minicpm-v	Coding + vision local
24 GB	gemma2:27b	llama3.2-vision	Strong text + best vision
48 GB	llama3.3:70b	llama3.2-vision	Full power local
96+ GB	Multiple 70B models	llama3.2-vision	Full local council, $0

Setup auto-detects your VRAM and recommends models that fit concurrently. No root/sudo needed — Ollama installs to ~/.nvh/. Full GPU guide

---

Why nvHive

Council scored 68% higher than a single model — at $0 cost. Three free providers running in parallel outperformed a single model on accuracy, completeness, and coherence. Benchmark details below.

• Smart team assembly. nvHive generates expert agents for your question and matches each to the best LLM for their specialty — a "Security Engineer" agent routes to a security-strong provider, a "Database Expert" to one suited for database queries.
• Automatic orchestration. Coding tasks get a planner + coder + reviewer. Complex questions get a council. Simple questions get the fastest advisor. All automatic.
• Scales with what you have. 1 provider → single-model answers. 3+ providers → council on complex questions. Local GPU → free inference alongside cloud. DGX Spark → three 70B models in parallel, fully local.
• 4-layer safety guardrails. Command blocklist, filesystem boundary enforcement, secrets redaction, and resource limits.

---

Architecture

9 layers from pip install to GPU inference — install, setup, 4 user interfaces, intent detection, 5 execution modes, smart routing, tool registry, 23+ AI providers, and the hardware stack. Local-first with cloud fallback. Architecture docs

---

Features

Desktop Agent

AI that sees your screen, controls mouse/keyboard, installs software, and navigates browsers — powered by local vision models.

nvh "take a screenshot and describe my desktop"
nvh "setup comfyui"                    # agent: git clone → pip install → launch → verify
nvh "open firefox and go to github.com"

Vision pipeline: screenshot → local vision model (llama3.2-vision / minicpm-v) → coordinate estimation → action → verify. Falls back to cloud vision if no local model. Works on Linux (X11), macOS, and Windows. Desktop agent docs

Agentic Coding

Multi-model coding agent with dynamic expert referral, iterative QA, parallel execution, and vision/browser tools.

nvh agent "Fix the streaming timeout bug in council.py"
nvh agent "Add unit tests for auth" --dir ./myproject
nvh agent "Build the notification service" --sandbox     # Docker-isolated
nvh review                     # multi-model code review
nvh test-gen nvh/core/council.py     # AI test generation

Key capabilities: dynamic expert referral, iterative QA refinement, parallel pipeline, Docker sandbox, execution checkpoints with rollback, LLM drift detection, multi-repo workspaces, and VS Code extension. Scales from no-GPU (fully cloud) to DGX Spark (3 local 70B models). Agentic coding docs

Council Mode

Run the same query through multiple providers in parallel, then synthesize. Expert personas generated per query, each assigned to a different model. Responses analyzed for agreement, synthesized by a non-member provider with a confidence score.

nvh convene "Should we use Redis or Postgres for sessions?"   # 3 models → synthesis
nvh throwdown "Review this architecture for scalability"      # 3-pass deep analysis with critique

Different models have different blind spots — council surfaces all perspectives. Council with 3 free providers costs $0. Council docs

Smart Routing

Each request is scored across capability (40%), cost (30%), latency (20%), and health (10%), then routed to the highest-scoring provider. Routing improves over time — after 20 queries per provider, it's fully data-driven.

nvh ask --escalate "Design a distributed lock manager"    # try free first, upgrade if uncertain
nvh ask --verify "Is eval() safe in Python?"              # cross-model verification
nvh routing-stats    # see learned vs static scores
nvh health           # provider resilience dashboard

Local-first with NVIDIA GPUs: simple queries route to your GPU via Ollama — no cloud, no cost, no data leaving your machine. --prefer-nvidia gives a 1.3x routing bonus to NVIDIA hardware. Routing docs

---

Providers

23 providers. 63 models. 25 free — no credit card required.

Tier	Providers	Rate Limits
Free (no signup)	Ollama (local), LLM7	Unlimited / 30 RPM
Free (email signup)	Groq, GitHub Models, Cerebras, SambaNova, Cohere, AI21, SiliconFlow, HuggingFace	15-30 RPM
Free (account)	Google Gemini, Mistral, NVIDIA NIM	15-1000 RPM
Paid	OpenAI, Anthropic, DeepSeek, Fireworks, Together, OpenRouter, Grok	Pay per token

Full provider guide

---

Integrations

nvHive exposes a CLI (nvh), web dashboard (nvh webui), Python SDK (import nvh), MCP server for Claude Code, and OpenAI/Anthropic-compatible API proxies.

import nvh

response = await nvh.complete([{"role": "user", "content": "Explain quicksort"}])
result = await nvh.convene("Architecture review", cabinet="engineering")

Integration	Setup
Anthropic SDK	ANTHROPIC_BASE_URL=http://localhost:8000/v1/anthropic
OpenAI SDK	OPENAI_BASE_URL=http://localhost:8000/v1/proxy
Claude Code	claude mcp add nvhive -- python -m nvh.mcp_server
NemoClaw	nvh nemoclaw --start — NemoClaw docs

SDK & API reference | Claude Code integration | OpenClaw migration

---

Benchmark Results

Real data from NVIDIA DGX Spark (GB10, 120GB). 16 prompts across code generation, debugging, reasoning, math, creative writing, and Q&A. Judged by OpenAI with ground truth verification.

Mode	Accuracy	Completeness	Coherence	Overall	Cost
Single Model (Nemotron Super)	5.5	5.7	5.0	5.1	$0.00
Council (Ollama + Groq + Google)	9.0	8.0	9.0	8.6	$0.00

nvh bench              # GPU speed (tokens/sec)
nvh bench -q           # speed + quality comparison
nvh health             # provider resilience

Results vary by hardware and workload — run nvh bench to measure on your setup.

---

Core Commands

Command	What It Does
nvh "question"	Smart route to best available model
nvh convene "question"	Council consensus (3+ models)
nvh throwdown "question"	Three-pass deep analysis with critique
nvh agent "task"	Agentic coding with expert referral + QA
nvh review	Multi-model code review
nvh test-gen file.py	AI test generation with verification
nvh safe "question"	Local only — nothing leaves your machine
nvh serve	Start API server (OpenAI + Anthropic proxy)
nvh webui	Launch web dashboard
nvh health	Provider resilience dashboard
nvh bench	GPU speed test (tokens/sec)
nvh setup	Interactive provider setup
nvh doctor	Full diagnostic dump

Full command reference (50+ commands)

---

Documentation

Guide	Description
Getting Started	First-time setup
Commands	Full CLI reference (50+ commands)
Providers	23 providers, rate limits, free tiers
Council System	Multi-LLM consensus with confidence scoring
Architecture	System design and adaptive routing
GPU Detection	Auto-detection, model selection, OOM protection
SDK & API	Python SDK, REST API, proxies
Agent Tools	Agent tools and capabilities
Configuration	Configuration reference
Web UI	Web dashboard
Deploy Without Root	No-root install on servers
Windows Troubleshooting	Encoding, segfaults, port issues
Releasing	Release runbook

---

Important Notes

• Data Privacy: Cloud providers transmit queries to third-party APIs subject to each provider's privacy policy. Use nvh safe or --prefer-nvidia to keep inference local.
• AI Accuracy: AI-generated outputs may contain errors. Review agent-modified files before committing to production.
• Security: Safety guardrails use pattern-matching heuristics. For sensitive environments, use --sandbox with Docker isolation.
• Benchmarks: Results measured on NVIDIA DGX Spark reference hardware. Results vary by hardware, provider, and workload.

License

MIT License. See LICENSE for details.