modelmux 0.20.0

Model multiplexer — unified MCP server for cross-platform multi-model AI collaboration

modelmux

Model multiplexer — unified MCP server for cross-platform multi-model AI collaboration.

Route tasks to Codex CLI, Gemini CLI, and Claude Code CLI through a single MCP interface with smart routing and caller auto-detection.

Install

# One-command install for Claude Code
claude mcp add modelmux -s user -- uvx modelmux

# Codex CLI (~/.codex/config.toml)
# [mcp_servers.modelmux]
# command = "uvx"
# args = ["modelmux"]

# Gemini CLI (~/.gemini/settings.json)
# {"mcpServers": {"modelmux": {"command": "uvx", "args": ["modelmux"]}}}

Tools

• mux_dispatch — Send a task to a model and get structured results
  • provider: "auto" / "codex" / "gemini" / "claude" / "ollama"
  • task: The prompt to send
  • workdir, sandbox, session_id, timeout, model, profile, reasoning_effort, failover
• mux_check — Check which CLIs are available, show detected caller and config

Smart Routing

provider="auto" routes tasks by keyword analysis and auto-excludes the calling platform:

From Claude Code → routes to Codex or Gemini (never back to Claude)
From Codex CLI → routes to Claude or Gemini (never back to Codex)

Failover

When a provider fails at execution time, modelmux automatically retries with the next available provider (enabled by default, failover=True). Failover is skipped for session-based requests since sessions are provider-specific.

Audit & Policy

Every dispatch call is logged to ~/.config/modelmux/audit.jsonl for debugging and cost tracking.

Policy enforcement via ~/.config/modelmux/policy.json:

{
  "blocked_providers": ["gemini"],
  "blocked_sandboxes": ["full"],
  "max_timeout": 600,
  "max_calls_per_hour": 30,
  "max_calls_per_day": 200
}

mux_check() now shows policy summary and audit stats.

User Configuration

Create .modelmux/profiles.toml or ~/.config/modelmux/profiles.toml:

[routing]
default_provider = "codex"

[[routing.rules]]
provider = "gemini"
[routing.rules.match]
keywords = ["frontend", "react", "css"]

[profiles.budget]
[profiles.budget.providers.codex]
model = "gpt-4.1-mini"

Links

• Full Documentation
• 中文文档

License

MIT