Multi Model

multi-model plugin

Portable Claude Code plugin for multi-model orchestration: delegate to Ollama cloud, NVIDIA NIM, NVIDIA Security, GitHub Copilot CLI, Google Gemini CLI, opencode CLI, and Codex from any Claude Code project.

Install

Add to your project's .claude-plugin/marketplace.json:

{
  "plugins": [
    { "name": "multi-model", "source": "./plugins/multi-model", "version": "1.3.0" }
  ]
}

Or reference this repo directly as a marketplace source.

Requirements

Dependency	Notes
node on PATH	Required for all MCP server scripts
npm install inside the plugin	Installs @modelcontextprotocol/sdk, zod, js-yaml into local node_modules. No global install required.
OLLAMA_HOST	Optional. Default: http://localhost:11434. Must be running for Ollama cloud models.
NVIDIA_API_KEY	Required for NVIDIA NIM and NVIDIA Security commands
GH_TOKEN or GITHUB_TOKEN	Required for GitHub Copilot CLI; PAT with Copilot Requests scope
npm install -g @github/copilot	Required for Copilot CLI integration
GEMINI_API_KEY or Google OAuth	Required for Gemini CLI. Run gemini once interactively for OAuth, or set GEMINI_API_KEY.
npm install -g opencode-ai	Required for opencode CLI. Auth: opencode providers login.
npm install -g @openai/codex	Required for direct codex CLI (preferred Codex path). Bypasses openai-codex plugin's Landlock sandbox.
codex login (ChatGPT subscription)	Auth source. Credentials cached in ~/.codex/auth.json. No OPENAI_API_KEY needed.
openai-codex plugin installed	Optional fallback for /codex:* slash commands when the direct CLI is unavailable.
MCP_GLOBAL_MODULES	Optional (legacy). Fallback path to a global node_modules if plugin-local node_modules is missing.

Commands

Command	Description
/multi-model:delegate	Auto-route a task — Opus picks models, dispatches in parallel, Codex verifies
/multi-model:ollama	Delegate a prompt to an Ollama cloud model
/multi-model:nvidia	Delegate a prompt to a NVIDIA NIM frontier model
/multi-model:nvidia-security	Security audit / PII / guardrail task via NVIDIA Security NIM
/multi-model:copilot	Delegate a prompt to GitHub Copilot CLI (1 premium request per call)
/multi-model:gemini	Delegate a prompt to Google Gemini CLI (no premium cost)
/multi-model:opencode	Delegate a prompt to opencode CLI (free-tier models only)
/multi-model:codex	Hand off to Codex for review, rescue, or adversarial verification
/multi-model:models	List all available delegation models across providers
/multi-model:test-all-models	Fire canary ping against every executor and MCP endpoint, render status table

Skills

Skill	Trigger
multi-model-orchestrator	Auto-triggers on any non-trivial task. Opus plans + synthesizes; Sonnet/Haiku/Ollama/NVIDIA/Copilot/Gemini/Codex execute in parallel; Codex verifies.
multi-model-advisor	Triggered by questions about models — "which model for X?", "compare A vs B", "recommend", "tradeoffs". Does not dispatch work.
multi-model-ollama-models	Triggered by "ollama", "gemma", "kimi", "deepseek", "devstral", "glm", "qwen3-coder", second-opinion, vision/multimodal, deep CoT, or long-context agentic work.
multi-model-nvidia-nim-models	Triggered by "nvidia", "nemotron", "qwen3-coder", "devstral", "gemma4", "llama405b", or frontier coding/reasoning needs. Requires NVIDIA_API_KEY.
multi-model-nvidia-security-models	Triggered for security audits, CVE/OWASP/SAST triage, threat modeling, prompt-injection detection, PII scrubbing, compliance review. Requires NVIDIA_API_KEY.
multi-model-copilot-models	Triggered by "copilot", "gpt-5.3-codex", "gemini-3-pro", or cross-vendor second-opinion requests. Requires GITHUB_TOKEN.
multi-model-gemini-cli-models	Triggered by "gemini", Google-stack tasks, long-context, or multimodal work via Google's own CLI. No premium cost.

MCP Servers

Seven MCP servers are bundled and auto-loaded via plugin.json. No project-level .mcp.json needed.

Server	Script	Tools provided
ollama	scripts/mcp-ollama.mjs	ollama_chat, ollama_list_models — 15 Ollama cloud models
nvidia-nim	scripts/mcp-nvidia.mjs	nvidia_chat, nvidia_list_models — 10 NVIDIA NIM frontier models
nvidia-security	scripts/mcp-security-nvidia.mjs	nvidia_security_chat, nvidia_security_list_models — 8 security/audit/guardrail models
copilot	scripts/mcp-copilot.mjs	copilot_chat, copilot_list_models — 6 cross-vendor models via GitHub Copilot
gemini	scripts/mcp-gemini.mjs	gemini_chat, gemini_list_models — 5 Google Gemini models (no premium cost)
opencode	scripts/mcp-opencode.mjs	opencode_run, opencode_list_models — 4 free-tier cross-vendor models
codex	scripts/mcp-codex.mjs	codex_exec, codex_review — direct codex exec CLI (bypasses openai-codex plugin sandbox)

Configuration

Environment variables

Variable	Required	Description
NVIDIA_API_KEY	For NVIDIA commands	API key from build.nvidia.com
OLLAMA_HOST	Optional	Ollama server URL. Default: http://localhost:11434
GH_TOKEN / GITHUB_TOKEN	For Copilot command	GitHub PAT with Copilot Requests scope
GEMINI_API_KEY	For Gemini command	Google AI API key. Alt: GOOGLE_API_KEY, GOOGLE_APPLICATION_CREDENTIALS, or Google OAuth.
OPENCODE_SERVER_PASSWORD	Optional	opencode server auth. Alt: opencode providers login for interactive auth.
MCP_GLOBAL_MODULES	Optional (legacy)	Fallback path to a global node_modules. Prefer npm install inside the plugin.

Models

44 models across 6 providers (Ollama cloud, NVIDIA NIM, NVIDIA Security, GitHub Copilot CLI, Google Gemini CLI) plus 4 free-tier opencode models. See MODELS.md for the full catalog, per-model notes, and the recommended picks per lane (agentic-coding, deep-reasoning, multimodal, long-context, bulk, security-audit, pii, safety-guard, cross-vendor).

The catalog is generated from models.yaml. To regenerate:

cd plugins/multi-model
npm run gen-docs

Cost

Provider	Cost
Copilot CLI	1 premium request per call
Gemini CLI	Free (user's Google account quota)
opencode CLI	Free (free-tier models only)
Ollama cloud	Self-hosted / cloud endpoint
NVIDIA NIM	Per NVIDIA API pricing
Codex	Per OpenAI/ChatGPT subscription