Nexus Agents

May 15, 2026 · View on GitHub

Governance substrate for your AI coding agents — adversarial review, drift-detected rules, immutable audit, closed-loop telemetry

Why Nexus Agents?

Nexus-agents is a governance layer that sits above your AI coding agents — Claude Code, Codex, Gemini, and OpenCode. The agents do the engineering; nexus-agents enforces the rules they have to follow, reviews their work adversarially before it ships, audits everything they touch, and routes the next task based on what actually worked.

What it gives you:

Adversarial PR review — pr_review runs 5 voter roles (architect, security, devex, catfish, scope_steward) with a 4-point verification gate. On the v5 evaluation set (10 PRs): 100% bug-catch and 50% raw false-positive rate; manual triage reclassified most "FPs" as legitimate findings the dataset had mislabeled. Full numbers: docs/research/pr-review-experiment-results-v5.md
Drift-detected charter — CLAUDE.md + governance:check + blocking CI gates fail the build when documented rules drift from registered behavior (model registry, MCP tools, expert types, skills)
Immutable audit trail — every tool call, every voter decision, every routing choice flows through AuditTrail with structured logging and hash-chained append-only storage; integrity is verifiable via the verify_audit_chain MCP tool
Closed-loop routing — OutcomeStore feeds production telemetry back into LinUCB + TOPSIS scoring so the system actually learns from what shipped vs what regressed
Multi-voter consensus — consensus_vote runs a default 7-role panel (architect, security, devex, ai_ml, pm, catfish, scope_steward; --quick uses 3). Six strategies: simple/super-majority, unanimous, higher-order Bayesian, opinion-wise, proof-of-learning

You:               "Review this PR / orchestrate this task / vote on this proposal"
                    ↓
nexus-agents:       enforce rules → route → adversarial review → audit → learn from outcome
                    ↓
Engineering agents: Claude Code · Codex · Gemini · OpenCode
                    ↓
Code:               actual edits, tests, PRs, issues

What this is NOT:

Not another autonomous coding agent. OpenHands, SWE-agent, AutoGen, Devin, Factory — those are agents. Nexus-agents is the layer above them. Use whichever agents fit; we govern them
Not a chat framework. Nothing here orchestrates conversations. It orchestrates real CLI tool invocations with real file I/O and outcome tracking
Not a model API proxy. The value is the rules, the gates, the audit, and the learning. Routing is a side effect of the governance work, not the product

Where nexus-agents sits in your stack

   Human / IDE / CLI
   (Claude Code, Cursor, VS Code, terminal)
            │ MCP Protocol
            ▼
  ┌─────────────────────────────────────────────────────┐
  │  GOVERNANCE SUBSTRATE — what nexus-agents provides   │
  │                                                       │
  │   Charter (drift-checked)   Adversarial PR review    │
  │   Role registry             Multi-voter consensus    │
  │   Immutable audit trail     Closed-loop telemetry    │
  │                                                       │
  │   38 MCP tools · multi-stage CompositeRouter         │
  └────────────────────────┬────────────────────────────┘
                           │
                           ▼ delegates execution to
  ┌─────────────────────────────────────────────────────┐
  │  ENGINEERING AGENTS — what does the actual work      │
  │                                                       │
  │   Claude Code · Codex · Gemini · OpenCode            │
  └────────────────────────┬────────────────────────────┘
                           │
                           ▼ produces
                   Code, tests, PRs, issues

The governance substrate is the layer that catches the mistakes engineering agents would otherwise make — bad code shipped, rules drifting from intent, audit gaps, telemetry-free routing — and routes the next task based on what actually worked the last time.

Quick Start (2 minutes)

1. Install

npm install -g nexus-agents

Or as a Claude Code plugin (single-command install from the official marketplace):

/plugin install nexus-agents

See docs/getting-started/PLUGIN_INSTALL.md for plugin-specific setup, or llms-install.md for the short install guide an AI agent can follow.

2. Verify

nexus-agents doctor

3. Use

With Claude Code (recommended):

nexus-agents setup   # Auto-configures MCP server

Then in Claude: "orchestrate: Review this PR for issues"

Standalone CLI:

export ANTHROPIC_API_KEY=your-key
nexus-agents orchestrate "Explain the architecture of this codebase"

Security: In default MCP mode, the server communicates only via stdio with the parent process (no network exposure). The REST API (opt-in) auto-generates an API key on first start. For network-exposed deployments, set NEXUS_AUTH_ENABLED=true. See SECURITY.md.

Capabilities

Category	Details
Adversarial PR Review	`pr_review` MCP tool: 5 voter roles (architect, security, devex, catfish, scope_steward) with 4-point gate. v5 evaluation (n=10): 100% bug-catch, 50% raw FP rate; manual triage reclassified most FPs as legitimate findings (details)
Consensus Voting	6 strategies: simple_majority, supermajority, unanimous, higher_order (Bayesian correlation-aware), opinion_wise, proof_of_learning
Drift-Detected Charter	`CLAUDE.md` + `inject-governance.ts check` enforces single-source registries (model registry, MCP tools, expert types). Blocking CI gate fails build on drift
Audit Trail	Structured logging for every tool call, voter decision, and routing choice. Hash-chained immutable storage; integrity verifiable via `verify_audit_chain` MCP tool (#2281, #2289)
Closed-Loop Telemetry	`OutcomeStore` records production results; LinUCB bandit + TOPSIS scoring + adaptive routing bonuses adjust based on what actually worked. No other framework closes this loop
Security Pipeline	Sandboxing (Docker/policy), trust-tiered input handling, SARIF parsing, red-team patterns, ClawGuard access policies (audit/enforce)
Multi-Expert Orchestration	12 built-in expert types coordinated by Orchestrator. Roles bind prompt + tools + memory
Development Pipeline	Research → Plan → Vote → Decompose → Implement → QA → Security. Three modes: autonomous, harness (caller implements), dry-run
Memory & Learning	5 user-facing backends (session, belief, agentic, adaptive, typed). Cross-session persistence feeds routing decisions
Research System	9 discovery sources (arXiv, GitHub, Semantic Scholar, etc). Auto-catalog, quality scoring, synthesis into topic clusters
Graph Workflows	DAG-based workflow execution with checkpoint/resume, state reduction, and event hooks
38 MCP Tools	Agent management, workflow execution, research, memory, codebase intelligence, repo analysis, consensus, operations

Available Experts

Expert	Specialization
Code	Implementation, debugging, optimization
Architecture	System design, patterns, scalability
Security	Vulnerability analysis, secure coding
Testing	Test strategies, coverage, test generation
QA	Acceptance criteria, regression checks
Documentation	Technical writing, API docs
DevOps	CI/CD, deployment, infrastructure
Research	Literature review, state-of-the-art analysis
PM	Product management, requirements, priorities
UX	User experience, usability, accessibility
Infrastructure	Server management, bare metal, networking
Data Viz	Charts, dashboards, visual data presentation

Supported CLIs & Providers

Nexus-agents routes tasks through 5 CLI adapters, each connecting to major AI providers:

CLI	Provider	Best For
claude	Anthropic (Claude)	Complex reasoning, analysis
gemini	Google (Gemini)	Long context, multimodal
codex	OpenAI (Codex CLI)	Code generation, reasoning
codex-mcp	OpenAI (Codex MCP)	MCP-native Codex integration
opencode	Custom OpenAI-compat	Custom endpoints, local models

CLI Commands

nexus-agents                    # Start MCP server (default)
nexus-agents doctor             # Check installation health
nexus-agents setup              # Configure Claude CLI integration
nexus-agents orchestrate "..."  # Run task with experts
nexus-agents vote "proposal"    # Multi-agent consensus voting
nexus-agents review <pr-url>    # Review a GitHub PR
nexus-agents expert list        # List available experts
nexus-agents workflow list      # List workflow templates
nexus-agents config init        # Generate config file
nexus-agents init --portable    # Create workspace-local .nexus-agents/ for sandboxes
nexus-agents init --portable --mcp-config  # Also emit .mcp.json wiring Claude Code to it
nexus-agents init --portable --install --mcp-config  # …and install the binary into the workspace
nexus-agents fitness-audit      # Run fitness score audit
nexus-agents research query     # Query research registry
nexus-agents --help             # Full command list

See docs/ENTRYPOINTS.md for the complete CLI reference (28+ commands).

MCP Tools

When running as an MCP server, the following tools are available:

Tool	Description
`orchestrate`	Task orchestration with Orchestrator coordination
`create_expert`	Create a specialized expert agent
`execute_expert`	Run a task through a previously-created expert (by expertId)
`run_workflow`	Run a linear workflow template (use `run_graph_workflow` for DAGs)
`delegate_to_model`	Pick the best-fit existing model for a task (no registry change)
`list_experts`	Inventory of expert ROLES for `create_expert`
`list_workflows`	Inventory of multi-step TEMPLATES for `run_workflow`
`consensus_vote`	Multi-model consensus voting on proposals
`research_query`	Query research registry (status, overlap, stats, search)
`research_add`	Add an arXiv PAPER to the registry (for non-paper sources use `research_add_source`)
`research_add_source`	Add a NON-PAPER source (repo/tool/blog) — for arXiv papers use `research_add`
`research_discover`	Discover papers/repos from external sources
`research_analyze`	Analyze registry for gaps, trends, coverage
`research_catalog_review`	Review auto-cataloged research references
`research_synthesize`	Synthesize registry into topic clusters with themes
`survey_oss_landscape`	Transient OSS project search (license, stars, last-commit) via GitHub
`vendor_publishing_audit`	Look up a vendor's signing infrastructure (GPG keys, URL patterns, signature shape)
`compare_data_feeds`	Diff two YAML/JSON feeds: coverage + per-field axes
`memory_query`	Query across all memory backends
`memory_stats`	Memory system statistics dashboard
`memory_write`	Write to typed memory backends
`weather_report`	Multi-CLI performance weather report
`issue_triage`	Triage GitHub issues with trust classification
`run_graph_workflow`	Run a DAG workflow with checkpoint + rollback (linear → `run_workflow`)
`execute_spec`	Execute AI software factory spec pipeline
`registry_import`	Draft YAML for a NEW model entry (for picking existing models use `delegate_to_model`)
`query_trace`	Query execution traces for observability
`query_task_state`	Query the structured task-state log for a task ID
`verify_audit_chain`	Verify hash chain of a FileAuditStorage audit log directory
`repo_analyze`	Analyze GitHub repository structure
`repo_security_plan`	Generate security scanning pipeline for a repo
`extract_symbols`	Tree-sitter AST symbols from a SINGLE file (functions/classes/types)
`search_codebase`	Cross-file ripgrep search for patterns or text (not an AST parser)
`run_dev_pipeline`	Full dev pipeline: research, plan, vote, implement, QA
`run_pipeline`	Execute a pipeline plugin by name with typed input
`pr_review`	Multi-voter PR review with verification gate (experimental)
`supply_chain_tradeoff_panel`	Per-axis tradeoff vote for build-vs-buy / supply-chain decisions
`improvement_review`	Threshold-gated observability loop — surfaces routing/tech-debt/bug/security signals from outcome+fitness data; files candidate issues

Configuration

Environment Variables:

Variable	Description
`ANTHROPIC_API_KEY`	Claude API key
`OPENAI_API_KEY`	OpenAI API key
`GOOGLE_AI_API_KEY`	Gemini API key
`NEXUS_LOG_LEVEL`	Log level (debug/info/warn/error)

Generate config file:

nexus-agents config init   # Creates nexus-agents.yaml

Documentation

Topic	Link
Full CLI Reference	docs/ENTRYPOINTS.md
Architecture	docs/architecture/README.md
Contributing	CONTRIBUTING.md
Coding Standards	CODING_STANDARDS.md
Quick Start Guide	QUICK_START.md

Development

git clone https://github.com/williamzujkowski/nexus-agents.git
cd nexus-agents
pnpm install
pnpm build
pnpm test

Requirements: Node.js 22.x LTS, pnpm 9.x

Contributing

Fork the repository
Create a feature branch (git checkout -b feat/amazing-feature)
Commit with conventional commits (feat(scope): add feature)
Open a Pull Request

See CONTRIBUTING.md for details.

License

MIT - See LICENSE

Built with Claude Code