nexus-agents weekly downloads statistics

Nexus Agents

Governance substrate for your AI coding agents — adversarial review, drift-detected rules, tamper-evident append-only audit chain, human-gated closed-loop tuning (autonomous demotion, earned promotion)

Overview

Nexus Agents is a governance layer that sits above your AI coding agents — Claude, Codex, Gemini, and OpenCode. The agents do the engineering; nexus-agents enforces the rules they follow, reviews their work adversarially before it ships, audits everything they touch, and routes the next task based on what actually worked (LinUCB bandit + TOPSIS scoring fed by an outcome store). It connects to any MCP-compatible editor. Routing is a means to the governance end, not the product.

Key Capabilities

• Intelligent Routing — Multi-stage CompositeRouter (TOPSIS, LinUCB bandit, distilled-rule short-circuit, weather-aware penalties). Learns from outcomes. • Multi-Expert Orchestration — Specialized expert agents (code, architecture, security, testing, docs, devops, research, PM, UX, infrastructure, QA, data-visualization) coordinated by an Orchestrator. • Consensus Voting — Five aggregation strategies: simple/super-majority, unanimous, higher-order Bayesian, opinion-wise. • Development Pipeline — Research → Plan → Vote → Decompose → Implement → QA → Security. Autonomous, harness, and dry-run modes. • Memory & Learning — Multiple backends (session, belief, adaptive, routing, graph, hybrid, agentic, typed) with cross-session persistence. • MCP Tools — Agent management, workflow execution, research, memory, codebase intelligence, repo analysis, consensus, operations. See docs/ENTRYPOINTS.md for the canonical list. • Research System — Discovery across arXiv, GitHub, Semantic Scholar, and other sources with auto-catalog and synthesis. • Security — Sandboxing, trust classification, SARIF parsing, input sanitization, red team pipeline.

Quick Start

Installation

CLI Usage

Start the MCP server:

Programmatic Usage

Claude Desktop Integration

Add to your Claude Desktop configuration (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

Features

Multi-Agent Orchestration

The Orchestrator agent analyzes incoming tasks and delegates to specialized experts:

| Expert | Specialization |

| Code Expert | Implementation, debugging, optimization, refactoring | | Architecture Expert | System design, patterns, trade-offs, scalability | | Security Expert | Vulnerability analysis, secure coding, threat modeling | | Documentation Expert | Technical writing, API docs, code comments | | Testing Expert | Test strategies, coverage analysis, test generation | | DevOps Expert | CI/CD, deployment, containerization | | Research Expert | Literature review, state-of-the-art analysis | | PM Expert | Product management, requirements, priorities | | UX Expert | User experience, usability, accessibility | | Infrastructure Expert | Server management, bare metal, networking | | QA Expert | Quality gates, test-plan review, defect triage | | Data-Visualization Expert | Charts, diagrams, dashboards, data presentation |

All twelve expert types can collaborate on complex tasks. The Orchestrator combines their outputs into a single response.

Model Adapters

Use different AI models through unified interfaces:

| Provider | CLI | Best For |

| Claude | claude | Complex reasoning, analysis | | Gemini | gemini | Long context, multimodal | | Codex | codex | Code generation, reasoning | | OpenCode | opencode | Custom OpenAI-compatible endpoints |

Model selection uses composite routing: Budget → Zero-cost → Preference → TOPSIS → LinUCB bandit.

Workflow Engine

Define reusable workflows in YAML:

MCP Tools

The server exposes 47 MCP tools for integration. Key tools include:

| Tool | Description |

| orchestrate | Analyze task and coordinate expert execution | | create_expert | Create a specialized expert agent | | execute_expert | Execute a task using a created expert | | run_workflow | Execute a workflow template | | delegate_to_model | Route task to optimal model | | consensus_vote | Multi-model consensus voting on proposals | | research_discover | Discover papers/repos from external sources | | memory_query | Query across all memory backends | | issue_triage | Triage GitHub issues with trust classification | | repo_analyze | Analyze GitHub repository structure | | repo_security_plan | Generate security scanning pipeline for a repo |

See the root README for the complete tool list.

Architecture

See docs/architecture/README.md for detailed module descriptions.

Dependency Flow

Core Interfaces

Configuration

Environment Variables

| Variable | Description | Required |

| ANTHROPIC_API_KEY | Claude API key | Yes (for Claude) | | OPENAI_API_KEY | OpenAI API key | For OpenAI models | | GOOGLE_AI_API_KEY | Google AI API key | For Gemini models | | OLLAMA_HOST | Ollama server URL | For Ollama (default: localhost:11434) | | NEXUS_LOG_LEVEL | Log level (debug/info/warn/error) | No |

API Reference

Adapters

Agents

MCP Server

Development

Prerequisites

• Node.js 24.x LTS • pnpm 9.x • TypeScript 5.9+

Setup

Commands

Contributing

We welcome contributions! Please see our guidelines:

1. Fork the repository 2. Create a feature branch (git checkout -b feat/amazing-feature) 3. Commit with conventional commits (git commit -m 'feat(agents): add amazing feature') 4. Push to your branch (git push origin feat/amazing-feature) 5. Open a Pull Request

Code Standards

• Files must be under 400 lines • Functions must be under 50 lines • Coverage thresholds: 60% statements/functions/lines, 50% branches (vitest.config.ts) • All code must pass linting and type checking

See CODING_STANDARDS.md for detailed guidelines. See CONTRIBUTING.md for contribution workflow.

Commit Convention

We use Conventional Commits:

License

MIT - See LICENSE for details.

Built with Claude Code