Give Claude Code access to 63 models across 23 providers — with one command.
# Option 1: If nvHive is installed via pip
claude mcp add nvhive -- python -m nvh.mcp_server
# Option 2: If using uvx (no install needed)
claude mcp add nvhive -- uvx nvhive mcp
# Option 3: Direct entry point
claude mcp add nvhive -- nvhive-mcpThat's it. Claude Code now has access to nvHive's full toolkit.
Once registered, Claude Code can use these tools in any session:
| Tool | What It Does |
|---|---|
ask |
Smart-routed query across 63 models — nvHive picks the best provider for the task |
ask_safe |
Local-only inference via Ollama — nothing leaves your machine |
council |
Multi-model consensus — 3+ LLMs debate and synthesize an answer |
throwdown |
Two-pass deep analysis with cross-critique between models |
status |
System health: providers, GPU, budget |
list_advisors |
Available providers and their status |
list_cabinets |
Expert persona presets (engineering, security_review, executive, etc.) |
Claude Code uses Anthropic models by default. With nvHive:
- Simple queries route to free providers (Groq, GitHub Models, LLM7, local Nemotron)
- Only complex tasks hit premium models
- Council mode: 3 free models often match a single premium model's quality
Ask Claude Code to verify its own answer using nvHive's council:
"Use the nvhive council tool to get a second opinion on this architecture decision"
Route sensitive code through local models only:
"Use nvhive ask_safe to analyze this credentials file locally"
Convene a council of specialists:
"Use the nvhive council tool with cabinet 'security_review' to audit this code"
You: "Use nvhive council with the code_review cabinet to review the changes in src/auth.py"
Claude Code → nvHive council → 3 expert personas review → synthesized feedback
You: "Use nvhive ask to research best practices for database indexing"
Claude Code → nvHive smart router → routes to free Groq/DeepSeek → instant answer at $0.00
You: "Use nvhive throwdown to analyze the tradeoffs of microservices vs monolith for our app"
Claude Code → nvHive throwdown → 3 LLMs analyze → 3 LLMs cross-critique → synthesis
nvHive reads config from ~/.hive/config.yaml. Key settings:
# Provider API keys
providers:
groq:
api_key: "gsk_..." # Free tier: 30 req/min
github:
api_key: "ghp_..." # Free tier: 150 req/day
openai:
api_key: "sk-..." # Pay per use
# Budget controls
budget:
daily_limit_usd: 1.00 # Hard stop at $1/day
warn_threshold: 0.80 # Alert at 80%
# Routing preferences
routing:
prefer_free: true # Prioritize free providers
prefer_local: false # Prioritize local Ollama models
fallback_enabled: true # Auto-fallback on provider failureRun nvh setup for an interactive configuration wizard.
nvHive's MCP server is compatible with any MCP client:
| Platform | Registration Command |
|---|---|
| Claude Code | claude mcp add nvhive -- python -m nvh.mcp_server |
| Claude Desktop | Add to claude_desktop_config.json |
| Cursor | Add to MCP settings |
| OpenClaw | Add to openclaw.json |
nvHive is building a Claude Code Channel plugin that pushes real-time events (cost alerts, council results, provider status) directly into your Claude Code session. See CLAUDE_CODE_CHANNELS_INTEGRATION.md for the roadmap.
"MCP SDK not installed"
pip install "mcp[cli]""No advisors enabled"
nvh setup # Interactive configuration
nvh test --quick # Verify provider connectivityServer not responding
# Test the MCP server directly
python -m nvh.mcp_server
# Or check Claude Code's MCP status
# In Claude Code, type: /mcp