multi-model

claude-code-multi-model

Portable Claude Code plugin for automatic multi-model orchestration. Opus plans + synthesizes. Sonnet/Haiku/Ollama cloud/NVIDIA NIM/NVIDIA Security/Codex execute in parallel. Codex verifies before merge. No user prompting for model choice — Opus auto-routes from task signal.

Drop into any project and start delegating across providers immediately.

---

What you get

• One plugin (multi-model) bundling 3 MCP servers + 6 slash commands + 1 auto-trigger skill.
• Auto-routing: Opus picks the right model per task silently.
• Parallel execution: independent subtasks dispatched in one message.
• Verification gate: Codex reviews every non-trivial diff before done.
• Portable: zero project-level .mcp.json needed — the plugin manifest loads the MCP servers.

---

Requirements

Requirement	Notes
Claude Code	Version with plugin + marketplace support
Node.js ≥ 18	on PATH
@modelcontextprotocol/sdk, zod (global npm)	npm i -g @modelcontextprotocol/sdk zod
MCP_GLOBAL_MODULES env	Points at your global node_modules. Windows: C:\Users\<you>\AppData\Roaming\npm\node_modules. macOS/Linux: output of npm root -g.
NVIDIA_API_KEY (optional)	For NVIDIA NIM + Security. Get at build.nvidia.com.
OLLAMA_HOST (optional)	Default http://localhost:11434. Ollama cloud models require an Ollama install + cloud-enabled account.
Codex plugin (optional)	For /codex:review, /codex:rescue, /codex:adversarial-review. Install from openai/codex-plugin-cc. Requires the Codex CLI on PATH.

---

Install in any project

Option A — GitHub marketplace (recommended)

Two commands, any project, any machine:

claude plugin marketplace add ranjankumarpatel/claude-code-multi-model
claude plugin install multi-model@claude-code-multi-model

Restart Claude Code → plugin auto-loads with its 3 MCP servers. Verify:

claude mcp list        # expect plugin:multi-model:{ollama,nvidia-nim,nvidia-security}

Updates: claude plugin update multi-model@claude-code-multi-model.

Option B — local clone (development)

For hacking on the plugin itself:

git clone https://github.com/ranjankumarpatel/claude-code-multi-model.git
claude plugin marketplace add /absolute/path/to/claude-code-multi-model
claude plugin install multi-model@claude-code-multi-model

---

Environment setup

Set once per machine (shell profile):

# Required for MCP servers to find the SDK
export MCP_GLOBAL_MODULES="$(npm root -g)"

# Optional — NVIDIA NIM + Security
export NVIDIA_API_KEY="nvapi-..."

# Optional — override Ollama host
export OLLAMA_HOST="http://localhost:11434"

Windows PowerShell:

setx MCP_GLOBAL_MODULES "C:\Users\$env:USERNAME\AppData\Roaming\npm\node_modules"
setx NVIDIA_API_KEY "nvapi-..."

Install MCP deps globally:

npm i -g @modelcontextprotocol/sdk zod

---

Install Codex integration

Codex is optional but recommended — it's the verification gate + rescue executor in the auto-routing pattern.

1. Install the Codex CLI and sign in so codex runs on your terminal.
2. Install the Codex plugin (bundled in this marketplace):

   claude plugin install codex@claude-code-multi-model
   

3. Verify with /codex:review or /codex:rescue inside Claude Code.

If Codex is not installed, multi-model still works — auto-routing will simply skip the Codex verification step.

---

How auto-routing works

Opus never edits files or runs shell directly. It parses your request, decomposes into subtasks, and dispatches each to the best executor using this rubric:

Task signal	Auto-route to
Bulk read / grep / rename / format	Haiku
Multi-file refactor, debugging, tests	Sonnet
Deep chain-of-thought reasoning	kimi-k2-thinking:cloud or deepseek-r1
Coding second opinion / alt-frontier	gemma4:31b-cloud or nemotron-ultra
Long-context / agentic / vision	kimi-k2.5:cloud
Multilingual / non-English code	mistral-large
Large general-purpose	llama405b
Security audit / CVE / OWASP / PII / injection	NVIDIA Security
Stuck / failing tests / pre-merge verify	Codex
≥2 independent subtasks	Parallel in one message

You just state the goal. Opus reports the route in one line (e.g. Routing: refactor → Sonnet; rename → Haiku; audit → NVIDIA Security) and runs.

---

Slash commands