nvhive 0.27.0

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.

Why nvHive

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 "Review my staged changes"                   # → auto-detects review → multi-model review
nvh "Add tests for the auth module"              # → auto-detects test request → test generation
nvh "Should we use Redis or Postgres?"           # → auto-detects debate → council (3+ advisors)

You type one command. nvHive figures out the rest. It detects what you're asking for, checks which advisors are healthy, and assembles the best team for the task — automatically. More advisors connected = smarter behavior, with zero configuration.

What makes it different:

• Smart team assembly. nvHive doesn't just route to one model — it generates expert agents based on your question and matches each one to the best LLM for their specialty. A "Security Engineer" agent gets the LLM that scores highest on security tasks. A "Database Expert" gets the best at database queries. All based on real performance data from the learning engine.
• Automatic orchestration. Coding tasks get a planner + coder + reviewer. Complex questions get a council of specialists. Simple questions get the fastest advisor. All automatic based on intent detection and available advisors.
• Scales with what you have. 1 provider? Single-model answers. 3+ providers? Council automatically on complex questions, multi-model verification on code. Local GPU? Free inference alongside cloud. DGX Spark? Three 70B models in parallel, fully local.
• Performant by default. Uses all available advisors within reason. Simple questions don't trigger council. Budget limits always enforced. Switch to cost mode for minimal spend.
• 4-layer safety guardrails. Command blocklist, filesystem boundary enforcement, secrets redaction, and resource limits — the agent can't rm -rf / even with --yes.

flowchart LR
    QUERY[User Query] --> INTENT[Intent Detection<br/>coding · review · council]
    INTENT --> AGENTS[Generate Expert Agents<br/>based on topic keywords]

    AGENTS --> MATCH[Match Agents to LLMs<br/>learning engine scores]

    MATCH --> A1[Security Engineer<br/>→ Claude<br/>best at security]
    MATCH --> A2[Database Expert<br/>→ GPT-4o<br/>best at databases]
    MATCH --> A3[Backend Architect<br/>→ Llama 70B local<br/>strong coding]

    A1 --> SYNTH[Synthesize<br/>best of each expert]
    A2 --> SYNTH
    A3 --> SYNTH

    SYNTH --> RESULT[Unified Answer<br/>all perspectives integrated]

    style MATCH fill:#1a1a2e,color:#76B900,stroke:#76B900
    style SYNTH fill:#1a1a2e,color:#3b82f6,stroke:#3b82f6
    style RESULT fill:#76B900,color:#000

Architecture Overview

flowchart TB
    User[User Prompt] --> Intent[Intent Detection]
    Intent --> |simple| LLM[Single LLM Query]
    Intent --> |coding| Agent[Agent Assembly]
    Intent --> |iterative| Iter[Iterative Loop]
    Intent --> |review| Review[Code Review]
    Intent --> |testgen| TestGen[Test Generation]
    Agent --> Match[Match Agents to LLMs]
    Iter --> Match
    Match --> Pipeline[Parallel Pipeline]
    Pipeline --> Referrals[Recursive Referrals]
    Referrals --> QA{QA Gate}
    QA -->|PASSED| Output[Final Output]
    QA -->|FAILED| Feedback[Feed Back]
    Feedback --> Pipeline

---

Agentic Coding

Multi-model coding agent with recursive spawning, iterative QA convergence, parallel execution, and vision/browser tools. Scales from no-GPU to DGX Spark.

# One-time setup: pulls the right models for your GPU
nvh agent --setup

# Run a coding task
nvh agent "Fix the streaming timeout bug in council.py"
nvh agent "Add unit tests for the auth middleware" --dir ./myproject
nvh agent "Refactor the router to use health-aware selection" -y

# Advanced: sandbox, workspace, parallel pipeline
nvh agent "Build the notification service" --sandbox     # Docker-isolated execution
nvh agent "task" --workspace ./api,./frontend             # multi-repo context
nvh agent "refactor the auth module" --sandbox  # runs in Docker container

How it works: Intent detection classifies the task, the orchestrator generates expert agents matched to the best LLMs, agents run in parallel where possible, recursive referral spawning fills knowledge gaps on-demand, and an iterative QA loop refines until convergence.

flowchart LR
    TASK[Task] --> INTENT[Intent Detection<br/>13 task types]

    INTENT --> DECOMPOSE[Decompose<br/>into subtasks]

    DECOMPOSE --> MATCH[Match Agents → LLMs<br/>learning engine scores]

    MATCH --> PAR[Parallel Pipeline<br/>independent subtasks<br/>run concurrently]

    PAR --> RECURSE[Recursive Spawning<br/>agents request specialists<br/>on-demand via REFER:]

    RECURSE --> QA{Iterative QA<br/>PASSED?}

    QA -->|PARTIAL / FAILED| FEEDBACK[QA feedback<br/>→ new agents<br/>→ next round]
    FEEDBACK --> PAR

    QA -->|PASSED| DONE[Converged<br/>Files modified<br/>Git commit]

    style INTENT fill:#1a1a2e,color:#76B900,stroke:#76B900
    style PAR fill:#1a1a2e,color:#3b82f6,stroke:#3b82f6
    style RECURSE fill:#1a1a2e,color:#a855f7,stroke:#a855f7
    style QA fill:#1a1a2e,color:#f59e0b,stroke:#f59e0b
    style DONE fill:#76B900,color:#000

Iterative QA Loop

The iterative QA loop drives tasks to convergence through multiple rounds of generation, execution, and review. Enable with --iterative and control spend with budget_usd.

flowchart LR
    A[Task] --> B[Generate Agents]
    B --> C[Execute + Recursive Referrals]
    C --> D[Synthesize]
    D --> E{QA Review}
    E -->|PASSED| F[Done]
    E -->|PARTIAL/FAILED| G[Feed Feedback]
    G --> B
    E -->|Budget Exceeded| F

nvh agent "task" --iterative               # enable iterative QA convergence
nvh agent "task" --iterative --budget 2.50  # cap spend at $2.50

Key Capabilities

Feature	What It Does
Recursive Agent Spawning	Agents self-identify knowledge gaps and emit REFER: Need a Database Expert for sharding — the system dynamically spawns the specialist, gets the answer, and feeds it back. Max depth prevents infinite recursion.
Iterative QA Convergence	Generate agents → run with referrals → post-QA reviews → if gaps found, spawn new agents informed by feedback → repeat until PASSED or budget exhausted.
Parallel Pipeline	Decomposes tasks into independent subtasks, runs them concurrently (bounded semaphore), respects dependencies, VRAM-aware model swapping with context preservation.
Vision + Desktop Control	Screenshot capture, image analysis via vision LLMs (GPT-4o, Claude, Gemini, LLaVA), OCR, mouse/keyboard automation with pyautogui. Agents can see and interact with GUIs.
Browser Automation	Headless browser navigation, screenshots, form filling via Playwright. HTTP requests, process management, Docker tools.
Docker Sandbox	--sandbox flag runs agent shell commands inside a Docker container — memory-limited, CPU-limited, no network by default, non-root user. Falls back to local if Docker unavailable.
Execution Checkpoints	File state snapshots before execution. Automatic rollback on failure — restores modified files, deletes newly created ones.
LLM Drift Detection	Monitors provider quality over time using EMA. Alerts when a provider drops >20% vs historical average. Auto-reroutes traffic away from degraded providers.
Code Analysis	Static analysis for code smells (long functions, deep nesting, complex conditionals, magic numbers, missing docstrings), tech debt scoring, complexity hotspots, missing test detection.
Multi-Repo Workspaces	--workspace aggregates multiple repos into a single agent context. Cross-repo import detection, language detection, shared file patterns. Read-only support for reference repos.
VS Code Extension	Agent tasks, code review, test generation, council queries, and explain — all from the VS Code sidebar. Auto-starts nvh serve if needed.

Scales with your hardware — 6 tiers from no-GPU to DGX Spark:

GPU	VRAM	Tier	Models	Mode
DGX Spark	128 GB	Tier 5	Nemotron 70B + Llama 70B + Qwen 72B (3 models, all local)	Multi
RTX 6000 Pro BSE	96 GB	Tier 4	Cloud planner + Llama 70B coder + Qwen 32B reviewer (dual local)	Multi
A100 / A6000	48-80 GB	Tier 3	Cloud planner + Llama 70B coder (--mode multi for dual local)	Auto
RTX 3090 / 4090	24 GB	Tier 2	Cloud planner + Gemma 2 27B coder	Single
RTX 4060 Ti	16 GB	Tier 1	Cloud planner + Qwen Coder 7B	Single
No GPU	—	Tier 0	Fully cloud	Single

nvh agent --setup                    # pull recommended models
nvh agent --remove                   # clean up models
nvh agent "task" --mode multi        # force multi-model (Tier 3+)
nvh agent "task" --mode single       # force single model
nvh agent "task" --git               # auto-branch + commit changes
nvh agent "task" --no-quality        # skip lint/syntax gates

Multi-model mode (Tier 4-5, or --mode multi on Tier 3): a DIFFERENT model reviews the coder's output, catching bugs the coder's architecture has blind spots for. Cross-model verification is measurably better than self-review.

Quality gates: after the agent modifies files, ruff lint + syntax checks run automatically. If they fail, the agent gets the errors and fixes them in a feedback loop.

Code Review (nvh review)

nvh review                     # review staged changes
nvh review HEAD~3..HEAD        # review last 3 commits
nvh review 42                  # review GitHub PR #42
nvh review --mode multi        # two models review independently

Multi-model code review: two different LLM architectures review independently, then findings are synthesized. Catches bugs that self-review and single-model review miss.

Test Generation (nvh test-gen)

nvh test-gen nvh/core/council.py     # generate tests for a file
nvh test-gen --coverage-gaps         # find and fill coverage gaps

Reads your code, identifies untested paths, generates pytest tests, runs them, and iterates until they pass. The agent that improves itself — it writes the tests that verify its own future changes.

---

Get Started

pip install nvhive
nvh setup              # configure providers (validates keys)
nvh health             # check what's available
nvh "your question"    # try it

# With optional extras
pip install nvhive[all]      # vision + browser automation
pip install nvhive[vision]   # desktop control (pyautogui)
pip install nvhive[browser]  # browser automation (playwright)

First-Run Setup

On first run, nvh automatically launches guided setup:

• Detects GPU hardware and VRAM tier
• Shows provider status (which API keys are configured)
• Prompts for missing API keys with validation
• Offers to pull recommended Ollama models for your GPU

Works immediately with LLM7 (no signup). Run nvh setup to add free providers like Groq and GitHub Models.

NVIDIA GPU Quick Start — local inference on your hardware

# 1. Install Ollama + Nemotron
curl -fsSL https://ollama.com/install.sh | sh
ollama pull nemotron-mini        # 4.1GB, runs on 8GB+ VRAM

# 2. Install nvHive
pip install nvhive

# 3. nvHive auto-detects your GPU and Nemotron
nvh nvidia                       # GPU info + inference stack status
nvh bench                        # benchmark your GPU (tokens/sec)

# 4. Queries route to your GPU by default
nvh "Explain quicksort"          # → local Nemotron, $0, private
nvh safe "Analyze this code"     # → forced local, nothing leaves machine
nvh --prefer-nvidia "question"   # → 1.3x bonus for NVIDIA providers

# 5. Council on your GPU — 3 models, $0, fully private
nvh convene "Redis vs Postgres for sessions?"

nvHive detects NVIDIA GPUs via pynvml (VRAM, driver, CUDA version, temperature, power draw) and selects the optimal Nemotron model for your hardware. Simple queries stay local. Complex queries escalate to cloud only when needed. The learning loop measures your GPU's quality over time and adjusts routing thresholds automatically.

---

How It Works

Query Pipeline

flowchart TB
    USER[User Query] --> CLASSIFY[Task Classifier<br/>TF-IDF · 13 task types]
    CLASSIFY --> LOCALCHECK{Local GPU<br/>good enough?}
    
    LOCALCHECK -->|Simple query| GPU[NVIDIA GPU via Ollama<br/>Nemotron + Gemma 4<br/>Two architectures locally]
    LOCALCHECK -->|Complex query| SCORE[Score All Providers<br/>capability · cost · latency · health]
    
    SCORE --> ROUTE{Pick Best<br/>Provider}
    
    ROUTE --> FREE[Free Providers<br/>LLM7 · Groq · GitHub]
    ROUTE --> PAID[Premium Providers<br/>OpenAI · Anthropic · Google]
    ROUTE --> NIM[NVIDIA NIM<br/>Triton]
    ROUTE --> GPU
    
    FREE --> RESPONSE[Response]
    PAID --> RESPONSE
    NIM --> RESPONSE
    GPU --> RESPONSE
    
    RESPONSE --> LEARN[Learning Loop<br/>Record outcome · EMA update<br/>Adjusts GPU routing thresholds]
    LEARN -->|Feeds back into| SCORE
    
    RESPONSE -->|--verify flag| VERIFY[Cross-Model<br/>Verification]
    VERIFY --> FINAL[Verified Response]
    RESPONSE --> FINAL
    
    style GPU fill:#76B900,color:#000
    style NIM fill:#76B900,color:#000
    style LEARN fill:#1a1a2e,color:#76B900,stroke:#76B900
    style VERIFY fill:#1a1a2e,color:#00bcd4,stroke:#00bcd4

Task classification: TF-IDF cosine similarity against a 90-example training corpus (13 task types). Semantic understanding, not keyword matching.

Provider scoring: Weighted composite — capability (40%), cost (30%), latency (20%), health (10%). Capability scores start from static estimates and converge to measured performance via exponential moving average.

Adaptive learning: After every query, nvHive records the outcome and updates scores. By 20 queries per provider/task pair, routing is fully data-driven.

nvh routing-stats    # see learned vs static scores
nvh health           # provider resilience dashboard

Failover: If a provider fails, nvHive tries the next in the fallback chain. Every failure feeds back into the health score.

Local-first with NVIDIA GPUs: Simple queries route to Nemotron on your NVIDIA GPU via Ollama — no cloud, no cost, no data leaving your machine. GPU detection via pynvml reads VRAM, driver version, and CUDA version to select the optimal local model. The --prefer-nvidia flag gives a 1.3x routing bonus to keep inference on NVIDIA hardware whenever quality allows.

---

Council Mode

flowchart TB
    QUERY[User Query] --> AGENTS[Generate Expert Personas<br/>e.g. Backend Engineer, Architect, DBA]
    
    AGENTS --> M1[Model 1<br/>Groq / Llama]
    AGENTS --> M2[Model 2<br/>Google / Gemini]
    AGENTS --> M3[Model 3<br/>GitHub / GPT-4o]
    
    M1 --> COLLECT[Collect Responses<br/>Rate-limit staggered]
    M2 --> COLLECT
    M3 --> COLLECT
    
    COLLECT --> AGREE[Agreement Analysis<br/>Keyword overlap + LLM judge]
    AGREE --> SYNTH[Synthesis<br/>Uses non-member provider]
    
    SYNTH --> RESULT[Council Response<br/>+ Confidence Score<br/>+ Individual Perspectives]
    
    style AGREE fill:#1a1a2e,color:#00bcd4,stroke:#00bcd4
    style SYNTH fill:#1a1a2e,color:#76B900,stroke:#76B900

When one model isn't enough, nvHive runs the same query through multiple providers in parallel, then synthesizes their responses.

Why this works: Different models have different blind spots. Council mode surfaces all perspectives and synthesizes the best of each.

Confidence scoring: Every council response includes an agreement metric — "3/3 agreed" vs "split decision." Tells you when to trust the consensus.

Cost: Council with 3 free providers costs $0. Council with 3 Nemotron variants on a single NVIDIA GPU costs $0 and never leaves your machine. Premium cloud council costs ~3x a single query.

nvh convene "Should we use Redis or Postgres for session storage?"
# → 3 models debate → synthesis with confidence score

Throwdown Mode — Two-Pass Deep Analysis

Throwdown goes beyond council. Three passes, each building on the last:

flowchart TB
    QUERY[User Query] --> A1[Expert 1 - Nemotron<br/>local GPU]
    QUERY --> A2[Expert 2 - Gemma 4<br/>local GPU]
    QUERY --> A3[Expert 3 - Groq<br/>cloud free]
    
    A1 --> S1[Pass 1 Synthesis]
    A2 --> S1
    A3 --> S1
    
    S1 --> B1[Expert 1 - Critiques]
    S1 --> B2[Expert 2 - Finds blind spots]
    S1 --> B3[Expert 3 - Challenges assumptions]
    
    B1 --> S2[Pass 2 Synthesis]
    B2 --> S2
    B3 --> S2
    
    S2 --> FINAL[Final Answer]
    
    style A1 fill:#1a1a2e,stroke:#76B900,color:#c8c8c8
    style A2 fill:#1a1a2e,stroke:#76B900,color:#c8c8c8
    style A3 fill:#1a1a2e,stroke:#76B900,color:#c8c8c8
    style B1 fill:#1a1a2e,stroke:#00bcd4,color:#c8c8c8
    style B2 fill:#1a1a2e,stroke:#00bcd4,color:#c8c8c8
    style B3 fill:#1a1a2e,stroke:#00bcd4,color:#c8c8c8
    style FINAL fill:#76B900,color:#000

nvh throwdown "Review this architecture for scalability issues"
# Pass 1: 3 experts analyze independently
# Pass 2: experts critique each other's analysis
# Pass 3: final synthesis integrating all perspectives

Why throwdown beats single-model: A single model gives you one perspective, once. Throwdown gives you three perspectives, challenged by three critiques, then synthesized. Errors get caught. Assumptions get questioned. The final answer is more thorough than any single pass.

---

Smart Query Features

# Confidence-gated escalation: try free first, upgrade only if needed
nvh ask --escalate "Design a distributed lock manager"
# → groq (free, confidence: 42%) → auto-escalated to openai

# Cross-model verification: a second model checks the answer
nvh ask --verify "Is eval() safe in Python?"
# → groq answers → google verifies ✓ (9/10, no issues)

# Both together: cheapest possible verified answer
nvh ask --escalate --verify "Explain the CAP theorem"

---

Local GPU Inference with Nemotron

nvh setup detects your NVIDIA GPU, selects which models fit in your VRAM, and pulls them automatically. Supports both NVIDIA Nemotron and Google Gemma 4 (NVIDIA-optimized) for local council with two different architectures.

nvh setup
# Step 3/3: Local GPU inference
#   Detected: NVIDIA GeForce RTX 4090 (24GB VRAM)
#   Models: nemotron-small, gemma4:26b
#   Pulling nemotron-small... ✓
#   Pulling gemma4:26b... ✓
#   Local council ready — multiple models for consensus

What nvh setup handles:

flowchart TB
    SETUP[nvh setup] --> DETECT[GPU Detection<br/>pynvml reads VRAM · driver · CUDA]
    
    DETECT --> VRAM{Available VRAM?}
    
    VRAM -->|< 6 GB| MINI[nemotron-mini<br/>+ gemma4:e2b]
    VRAM -->|6 – 12 GB| SMALL[nemotron-small<br/>+ gemma4:e4b]
    VRAM -->|12 – 48 GB| CHOICE{User choice}
    VRAM -->|48 GB+| FULL[nemotron 70B<br/>+ gemma4:31b]

    CHOICE -->|Both for council| DUAL[nemotron-small<br/>+ gemma4:26b]
    CHOICE -->|Single model| SINGLE[nemotron 70B only]

    MINI --> CHECK{Ollama running?}
    SMALL --> CHECK
    DUAL --> CHECK
    SINGLE --> CHECK
    FULL --> CHECK
    
    CHECK -->|Not installed| INSTALL[Show install command]
    CHECK -->|Not running| START[Show: ollama serve]
    CHECK -->|Running| PULL[Auto-pull all<br/>models that fit]
    
    PULL --> READY[Ready ✓<br/>Local council enabled]
    
    READY --> ROUTE[nvHive Router<br/>Two model architectures<br/>Learning loop active]
    
    style SMALL fill:#76B900,color:#000
    style DUAL fill:#76B900,color:#000
    style READY fill:#76B900,color:#000
    style ROUTE fill:#1a1a2e,color:#76B900,stroke:#76B900

After setup, routing is automatic:

• Simple queries → local Nemotron or Gemma 4 on your GPU (free, private)
• Council mode → both models collaborate locally, catching different blind spots
• Complex queries → cloud providers when local quality isn't sufficient
• nvh bench measures your GPU's actual tok/s with community baselines
• The learning loop measures each model's quality on YOUR hardware

Full GPU detection + VRAM guide

NVIDIA Inference Stack

Layer	Technology	Hardware	Use Case
Local	Ollama + Nemotron	Consumer GPUs (RTX 3060+)	Default local inference, privacy mode
Local	Ollama + Gemma 4	Consumer GPUs (RTX 3060+)	NVIDIA-optimized, reasoning + multimodal
Cloud	NVIDIA NIM API	NVIDIA cloud	Specialized models, 1000 free credits
Enterprise	Triton Inference Server	H100 / A100 / L40	Production multi-model serving, TensorRT-LLM
Agent	NemoClaw / OpenShell	Any	Agent orchestration with nvHive routing
Detection	pynvml	Any NVIDIA GPU	VRAM, driver, CUDA, temp, power, PCIe

--prefer-nvidia gives a 1.3x routing bonus to all NVIDIA-backed providers, keeping inference on NVIDIA hardware whenever quality allows.

---

Integrations

How nvHive Connects to Your Tools

flowchart LR
    subgraph Your Tools
        CLI[nvh CLI<br/>agent · review · test-gen]
        WEBUI[Web Dashboard<br/>nvh webui]
        SDK[Python SDK<br/>import nvh]
        CC[Claude Code<br/>MCP]
        NC[NemoClaw<br/>Agent]
        CU[Cursor]
        APP[Your App<br/>OpenAI SDK]
    end

    subgraph nvHive Engine
        API[API Server<br/>:8000]
        MCP[MCP Server<br/>stdio]
        PROXY_OAI[OpenAI Proxy<br/>/v1/proxy]
        PROXY_ANT[Anthropic Proxy<br/>/v1/anthropic]
        AGENT[Agent Loop<br/>plan · execute · verify]
        ROUTER[Adaptive Router<br/>+ Learning Loop]
        COUNCIL[Council Engine<br/>+ Confidence]
        GUARD[Guardrails<br/>4-layer safety]
    end

    subgraph Providers
        GPU[Your GPU<br/>Ollama · Nemotron]
        FREE_P[Free Cloud<br/>Groq · GitHub · LLM7<br/>Google · Cerebras]
        PAID_P[Paid Cloud<br/>OpenAI · Anthropic<br/>DeepSeek · Mistral]
        NIM[NVIDIA NIM<br/>Triton]
    end

    CLI --> API
    WEBUI --> API
    SDK --> API
    CC --> MCP
    NC --> PROXY_OAI
    CU --> MCP
    APP --> PROXY_OAI
    APP --> PROXY_ANT

    MCP --> API
    PROXY_OAI --> API
    PROXY_ANT --> API
    API --> AGENT
    API --> ROUTER
    API --> COUNCIL
    AGENT --> GUARD
    GUARD --> ROUTER
    ROUTER --> GPU
    ROUTER --> FREE_P
    ROUTER --> PAID_P
    ROUTER --> NIM

    style GPU fill:#76B900,color:#000
    style NIM fill:#76B900,color:#000
    style ROUTER fill:#1a1a2e,color:#76B900,stroke:#76B900
    style COUNCIL fill:#1a1a2e,color:#00bcd4,stroke:#00bcd4
    style AGENT fill:#1a1a2e,color:#a855f7,stroke:#a855f7
    style GUARD fill:#1a1a2e,color:#ef4444,stroke:#ef4444

API Proxies — point existing SDKs at nvHive:

SDK	Configuration
Anthropic	ANTHROPIC_BASE_URL=http://localhost:8000/v1/anthropic
OpenAI	OPENAI_BASE_URL=http://localhost:8000/v1/proxy
Claude Code	claude mcp add nvhive -- python -m nvh.mcp_server
Cursor	nvh integrate --auto

Works With OpenClaw & NemoClaw

nvHive works alongside OpenClaw as a routing layer, and integrates with NemoClaw (NVIDIA's agent framework) as both inference provider and MCP tool server.

nvh migrate --from openclaw    # import your existing API keys
nvh nemoclaw --start           # start proxy for NemoClaw agents

Note: Anthropic recently changed billing for third-party tools. See the integration guide for details.

---

For Tool Builders

nvHive is a routing layer. Any AI application can add multi-provider routing:

import nvh

# Drop-in OpenAI-compatible interface
response = await nvh.complete([
    {"role": "user", "content": "Explain quicksort"}
])

# Inspect routing without executing
decision = await nvh.route("complex question about databases")

# Council consensus
result = await nvh.convene("Architecture review", cabinet="engineering")

# Provider health check
status = await nvh.health()

SDK & API reference

---

Core Commands

Agentic Coding

Command	What It Does
nvh agent "task"	Recursive agents + iterative QA convergence (6 GPU tiers)
nvh agent --setup	Pull recommended local models for your GPU
nvh agent --mode multi	Force multi-model: separate planner, coder, reviewer
nvh agent --sandbox	Execute shell commands inside a Docker container
nvh agent --workspace ./a,./b	Multi-repo context for cross-project tasks
nvh agent --git	Auto-create branch + commit changes
nvh review	Multi-model code review (staged changes, PRs, commit ranges)
nvh test-gen file.py	AI test generation with automatic verification
nvh analyze ./src	Code smells, tech debt score, complexity hotspots
nvh drift	Check for LLM quality degradation across providers

Queries & Council

Command	What It Does
nvh "question"	Smart route to best available model
nvh convene "question"	Council consensus (3+ models collaborate)
nvh throwdown "question"	Two-pass deep analysis with critique
nvh poll "question"	Side-by-side provider comparison
nvh safe "question"	Local only — nothing leaves your machine
nvh ask --escalate	Try free first, escalate if uncertain
nvh ask --verify	Cross-model verification

Infrastructure

Command	What It Does
nvh serve	Start the API server (OpenAI + Anthropic compatible proxy)
nvh webui	Launch the web dashboard
nvh health	Provider resilience dashboard
nvh nvidia	NVIDIA GPU infrastructure status
nvh bench	GPU speed test (tokens/sec)
nvh setup	Interactive provider setup
nvh doctor	Full diagnostic dump for troubleshooting

Full command reference (50+ commands)

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

---

Benchmark Results

Real data from NVIDIA DGX Spark (GB10, 120GB). Judged by OpenAI with ground truth verification on math prompts.

Quality: Council vs Single Model

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

Council consensus scored 68% higher than a single model on the same prompts. Ground truth verification on math problems caught errors the single model made that an LLM judge alone wouldn't have flagged.

Speed: Models on DGX Spark

Model	Size	tok/s
gemma3	3.3 GB	119.3
nemotron-mini	2.7 GB	85.7
gemma4 (e4b)	9.6 GB	61.7
llama3.1	4.9 GB	48.2
nemotron-3-super	86 GB	23.6

Run It Yourself

nvh bench              # GPU speed (tokens/sec)
nvh bench -q           # speed + quality comparison
nvh health             # provider resilience
nvh why                # explain last routing decision
nvh estimate --gpu rtx_4090   # predict tok/s on any GPU

16 prompts across code generation, debugging, reasoning, math, creative writing, and Q&A. LLM judge + ground truth verification. Run it yourself. Publish the results.

Learn More

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
Releasing	Release runbook, version bumps, PyPI publishing
Windows Troubleshooting	Encoding, segfaults, port 80, nvh.exe locks
GPU Detection	Auto-detection, model selection, OOM protection
Claude Code	MCP server setup
NemoClaw	NVIDIA NemoClaw integration
OpenClaw Integration	Works alongside OpenClaw
SDK & API	Python SDK, REST API, proxies
Deploy Without Root	No-root install on servers (Ollama, keyring, systemd user service)
Web UI	Web UI
Agent Tools	Agent tools
Configuration	Configuration
Architecture	System design and adaptive learning

License

MIT License. See LICENSE for details.