Open source · MIT · Node 22+

One MCP server. Every AI coding tool.

Nexus Agents routes each task to the best model across Claude, Codex, Gemini, and OpenCode — then remembers what worked. Plug it into Claude Code, Cursor, or any MCP client and your agent gets smarter with every run.

Get Started → View on GitHub ↗

38 MCP Tools

11 Expert Types

5 CLI Adapters

8 Memory Backends

12 Routing Stages

176+ Papers Tracked

How It Works

Every task flows through analysis, intelligent routing, quality validation, and outcome recording. The system learns from each execution and improves over time.

Intelligent Routing

A 12-stage pipeline analyzes your task and picks the best AI model. LinUCB contextual bandits learn from real outcomes — routing gets smarter with every task you run.

Dynamic Expert Creation

11 expert types are created on demand based on task analysis. Each gets a specialized system prompt, routed to the best CLI, and coordinated by TechLead and Orchestrator agents.

Consensus Quality Gates

5 voting algorithms — including higher-order Bayesian aggregation — validate outputs before they ship. Agents vote, disagree, and iterate until quality passes.

Continuous Learning

8 memory backends persist what works across sessions. Outcomes feed routing decisions, failed approaches are avoided, and successful patterns are reinforced automatically.

Development Pipeline

Full workflow: research papers, plan architecture, vote on proposals, decompose into tasks, implement in parallel, QA review with iteration, security scan. Three modes: autonomous, harness, and dry-run.

Research Discovery

176+ papers tracked across 9 sources (arXiv, GitHub, Semantic Scholar). New discoveries auto-trigger pipeline assessments. Quality scoring and topic synthesis built in.

The Learning Loop

This is the core differentiator. Every execution makes the next one better.

flowchart TD task["You give a task"] --> analysis["Task Analysis picks best model from 5 CLIs"] analysis --> expert["Expert executes 11 specialized types, created on demand"] expert --> consensus["Consensus validates multi-model voting catches errors"] consensus --> outcome["Outcome recorded success/failure feeds 8 memory backends"] outcome -.->|feeds next routing decision| analysis outcome --> better["Next task: routing is smarter, memory recalls what worked"] classDef primary fill:#6750A4,stroke:#6750A4,color:#ffffff; classDef accent fill:#e8def8,stroke:#6750A4,color:#1c1b1f; class task,better primary; class analysis,expert,consensus,outcome accent;

Quick Start

Install and connect to your editor in under 5 minutes.

1. Install

# Install globally
pnpm install -g nexus-agents

# Verify
nexus-agents doctor

2. Connect to Claude Code

# Auto-configure MCP server
nexus-agents setup

# Or manually: ~/.claude/mcp.json
{
  "mcpServers": {
    "nexus-agents": {
      "command": "nexus-agents",
      "args": ["--mode=server"]
    }
  }
}

Architecture

Tasks enter through MCP, route through a 12-stage pipeline, execute on the best available CLI, and get validated by consensus before outcomes feed back into the router.

flowchart TD mcp["MCP Protocol 38 tools · rate-limited · auth"] orch["Orchestrator task analysis · dynamic experts"] router["Composite Router · 12 stages Budget → Capability → Preference TOPSIS → LinUCB → Latency"] claude["Claude"] gemini["Gemini"] codex["Codex"] codexmcp["Codex-MCP"] opencode["OpenCode"] outcome[("Outcome Store feeds back")] subgraph services ["Supporting Services"] consensus["Consensus 5 algorithms"] memory["Memory 8 backends"] pipeline["Pipeline 3 modes"] security["Security sandbox"] end mcp --> orch orch --> router router --> claude router --> gemini router --> codex router --> codexmcp router --> opencode claude --> outcome gemini --> outcome codex --> outcome codexmcp --> outcome opencode --> outcome outcome -.->|learns routing| router orch -.-> services classDef primary fill:#6750A4,stroke:#6750A4,color:#ffffff; classDef cli fill:#e8def8,stroke:#6750A4,color:#1c1b1f; classDef store fill:#f5eff7,stroke:#6750A4,color:#1c1b1f,stroke-dasharray: 5 3; class mcp,orch,router primary; class claude,gemini,codex,codexmcp,opencode cli; class outcome store;

Documentation

Pick where to start based on what you want to do next.