HIVE
Home / _core / mcp / OpenRouter MCP — Setup Guide

OpenRouter MCP — Setup Guide

Edit on GitHub
On this page

HIVE ships with 7 multi-model skills that let you query any model on OpenRouter (Kimi, GPT-5, Grok, Gemini, Qwen, DeepSeek, and more) without leaving Claude Code.

#How it works

The MCP server (religa/multi_mcp) runs locally and exposes 5 tools to Claude Code:

Tool What it does
mcp__multi__chat Single model query
mcp__multi__codereview Code review with chosen model
mcp__multi__compare Same prompt to N models in parallel
mcp__multi__debate Multi-model consensus with cross-critique
mcp__multi__models List available aliases and IDs

The .mcp.json file at the HIVE root wires everything up — Claude Code loads it automatically when you open any HIVE directory.

#Setup (5 minutes)

#Step 1 — Get an OpenRouter API key

  1. Create a free account at openrouter.ai
  2. Go to Settings → Keys
  3. Create a key — copy it (sk-or-...)
  4. Add credits if you want to use paid models (free tier models are available, see _core/MODELS-MAP.md)

#Step 2 — Set the env variable

Mac / Linux — add to ~/.bashrc or ~/.zshrc:

export OPENROUTER_API_KEY="sk-or-your-key-here"

Then reload: source ~/.bashrc

Windows — add to PowerShell profile ($PROFILE):

$env:OPENROUTER_API_KEY = "sk-or-your-key-here"

Or set it permanently via System → Environment Variables.

#Step 3 — Run the install script

# Mac / Linux
export OPENROUTER_API_KEY="sk-or-..."
bash _core/mcp/install.sh

# Windows (Git Bash or WSL)
export OPENROUTER_API_KEY="sk-or-..."
bash _core/mcp/install.sh

The script:

  1. Validates Python 3.11+ and git
  2. Clones religa/multi_mcp into ~/.local/share/hive/multi_mcp
  3. Creates a virtualenv and installs dependencies
  4. Runs a smoke test with Kimi K2.6 (~$0.001)

#Step 4 — Test it

Open Claude Code in your HIVE directory:

claude
/second-opinion kimi "what is the capital of France?"

You should see Kimi's raw response inside a delimited block.

#The 7 multi-model skills

All skills are global — available in every squad context.

Skill Invocation Use when
Second opinion /second-opinion You want one model's perspective on a question
Consensus /consensus You want N models on the same question + synthesis
Anti-bubble /anti-bubble You want an outsider to read an asset (no company context)
ICP check /icp-check You want to simulate your ICP reading a piece of copy
DoD check /dod-check You want a cheap model to validate a PR against acceptance criteria
Cost compare /cost-compare You want to find the cheapest model that does the job for a worker
Code review model /code-review-model You want a code review from any OpenRouter model

All skills:

  • Ask which model to use if not passed as argument
  • Accept friendly aliases (kimi, gpt5.5, grok) or raw OpenRouter IDs
  • Deliver model responses raw (copy-paste, zero editing)
  • Apply the reasoning model max_tokens rule automatically (see MODELS-MAP)

#Model selection

Skills ask which model to use. Zero defaults — you always choose.

Quick reference for common use cases:

Use case Suggested models
Code review kimi, gpt5codex, qwencoder
Reasoning / analysis gpt5.5, grok, kimi-think
Marketing / copy feedback gemini-pro, gpt5.5, kimi
Anti-bubble (diverse corpus) gemini-pro, kimi, glm5.1
Cheap workers dsflash, qwenflash, granite
Free tier kimi-free, dsflash-free, qwencoder-free

Full catalog with prices, context windows, and benchmarks: _core/MODELS-MAP.md

#Troubleshooting

MCP "multi" not showing in Claude Code:

  • Check that .mcp.json exists in the HIVE root
  • Verify OPENROUTER_API_KEY is set: echo $OPENROUTER_API_KEY
  • Restart Claude Code

mcp__multi__chat returns an error:

  • Check OpenRouter credits: openrouter.ai/settings/credits
  • Verify the model ID is valid: curl -s https://openrouter.ai/api/v1/models | python3 -c "import json,sys; [print(m['id']) for m in json.load(sys.stdin)['data']]" | grep kimi

Reasoning model returns empty response:

  • This is the max_tokens gotcha — skills handle this automatically with retry
  • If it persists, pass --max-tokens 6000 explicitly

Re-run install after update:

bash _core/mcp/install.sh

The script is idempotent — safe to run multiple times.