nexus-agents MCP Server

LinUCB 밴딧을 사용하여 최적의 AI 모델(Claude, Codex, Gemini, OpenCode)에 작업을 라우팅하고, 합의 투표를 통해 검증하며, 결과로부터 학습하는 지능형 오케스트레이션 플랫폼입니다. 29개의 MCP 도구, 개발 파이프라인, 8개의 메모리 백엔드를 제공합니다.

GitHub

문서

Nexus Agents

Autonomic control plane for AI coding agents — one entry point, adversarial review, tamper-evident hash-chained audit, human-gated closed-loop tuning (autonomous demotion, earned promotion)

Why Nexus Agents?

Nexus-agents is an autonomic control plane for your AI coding agents — Claude Code, Codex, Gemini, and OpenCode. The agents are the data plane: they do the engineering. Nexus-agents is the control plane: it admits work through one entry point, reviews it adversarially before it ships, records every action in a tamper-evident event log, and closes the loop by tuning where the next task goes based on what actually worked.

Borrowing the vocabulary of autonomic computing: the system runs a MAPE-K loop — Monitor, Analyze, Plan, Execute over a shared Knowledge base — so that operating your agent fleet is, as much as the evidence allows, self-managing rather than hand-driven.

The control-plane mapping

Each classic control-plane role maps to a shipped nexus-agents component — the metaphor is load-bearing, not decoration:

Control-plane role	nexus-agents component	What it does
Scheduler	`run` / MetaOrchestrator	One entry point picks (and optionally runs) the right strategy for a goal
Admission control	gates (`pr_review`, `consensus_vote`, `run_quality_gate`)	Adversarial review and quality gates decide what is allowed to ship
Event log	`AuditTrail` hash chain + `verify_audit_chain`	Append-only, tamper-evident record of every decision
Data plane	engineering CLIs	Claude Code, Codex, Gemini, OpenCode do the file edits, tests, PRs

The MAPE-K loop

   ┌────────── Monitor ──────────┐        OutcomeStore · AuditTrail · swarm-health
   │                             ▼        adapter circuit-breaker signals
Execute ◀── Plan ◀── Analyze ◀───┘        LinUCB + TOPSIS scoring, consensus
   │         │                            MetaOrchestrator strategy choice
   │         └── route the next task ──────────────────────────────────────┐
   ▼                                                                        │
 run the strategy ── adversarial review ── audit ── feed outcome back ──────┘
                              shared Knowledge: OutcomeStore + memory backends + audit log

Self-* capabilities

Autonomic systems are described by their self-* properties. Each row below maps to a loop that exists in the codebase today — nothing here is aspirational, and the authority each loop carries is bounded by ADR-0017's authority ladder (observe → suggest → advisory → enforce):

Self-* property	What it means here	Implementing loop (shipped)
Self-configuring	Detects environment and wires itself in	`nexus-agents setup` / `doctor` (`cli-commands.ts`) — detects CLIs, writes MCP config, reports health
Self-healing	Routes around failing dependencies automatically	Adapter circuit-breaker + swarm-health demotion (`cli-adapters/circuit-breaker.ts`); a capped, auto-decaying, demotion-only `TuneAdjustmentStore` adjustment
Self-optimizing	Learns where the next task should go	Closed-loop `OutcomeStore` → LinUCB + TOPSIS scoring in the `CompositeRouter`
Self-protecting	Constrains what untrusted input and tools can do	Trust-tiered input handling, ClawGuard access policies (audit/enforce), Docker/policy sandboxing (`security/`)

Honesty note: these loops sit at different rungs of the authority ladder. The self-tuning demotion is enforce but bounded (capped, auto-decaying, demotion-only); learned selection and other promotions are still earned per-loop against an evidence threshold plus ratification, not flipped on by default. See ADR-0017.

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: 100% bug-catch on a focused synthetic dataset (n=10) and a 50% raw false-positive rate; manual triage reclassified one of two inspected FP cases as a real finding the dataset had mislabeled. These are directional small-n figures, not measured rates. Full numbers and guardrails: 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)
Tamper-evident 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 (tamper-evident, not tamper-proof — see the audit hash-chain threat model)
Closed-loop routing — OutcomeStore feeds production telemetry back into LinUCB + TOPSIS scoring so the system actually learns from what shipped vs what regressed. A second, bounded loop runs by default: a signal.swarm_unhealthy (adapter circuit-breaker / swarm-health) applies a small, capped, auto-decaying routing demotion via TuneAdjustmentStore — demotion-only, never zeroes a CLI, every adjustment audited, opt-out with NEXUS_TUNE_ENFORCE=false
Multi-voter consensus — consensus_vote runs a default 7-role panel (architect, security, devex, ai_ml, pm, catfish, scope_steward; --quick uses 3). Six strategy names (five distinct: higher_order is an alias of opinion_wise, #514): simple/super-majority, unanimous, higher-order Bayesian, opinion-wise, proof-of-learning

You:               "Review this PR / orchestrate this task / vote on this proposal"
                    ↓
Control plane:      admit → schedule/route → adversarial review → audit → learn from outcome
                    ↓
Data plane (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 the data plane. Nexus-agents is the control plane above them. Use whichever agents fit; we admit, review, audit, and route their work
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 admission gates, the audit, and the closed-loop tuning. Routing is a consequence of the control-plane work, not the product
Not fully autonomous. "Autonomic" means self-managing within bounds, not unsupervised. Every loop's authority is capped by the authority ladder (ADR-0017); promotions to higher authority are earned against evidence and human ratification, never flipped on by default

Where nexus-agents sits in your stack

   Human / IDE / CLI
   (Claude Code, Cursor, VS Code, terminal)
            │ MCP Protocol
            ▼
  ┌─────────────────────────────────────────────────────┐
  │  CONTROL PLANE — what nexus-agents provides          │
  │                                                       │
  │   Scheduler: run / MetaOrchestrator                  │
  │   Admission control: PR review · consensus · gates   │
  │   Event log: tamper-evident hash-chained audit       │
  │   Closed-loop self-tuning (MAPE-K)                   │
  │                                                       │
  │   46 MCP tools · multi-stage CompositeRouter         │
  └────────────────────────┬────────────────────────────┘
                           │
                           ▼ delegates execution to
  ┌─────────────────────────────────────────────────────┐
  │  DATA PLANE — the agents that do the actual work     │
  │                                                       │
  │   Claude Code · Codex · Gemini · OpenCode            │
  └────────────────────────┬────────────────────────────┘
                           │
                           ▼ produces
                   Code, tests, PRs, issues

The control plane is the layer that catches the mistakes data-plane 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

Prints a health table — Node version, configured CLIs (claude / codex / gemini / opencode), API keys missing vs present. Read-only; safe to run any time.

3. See what success looks like (60-second smoke task — no API keys needed)

nexus-agents vote --quick --proposal "Use SQLite over JSON files for the outcome store"

You should see:

Nexus Agents Consensus Vote
============================

Collecting votes from 3 agents (timeout: 60s each)...

Proposal: Use SQLite over JSON files for the outcome store

Votes

  ✓ Software Architect: APPROVE (86%)
  ✓ Security Engineer:  APPROVE (74%)
  ✓ Scope Steward:      APPROVE (91%)

Summary

  Approve:  3
  Reject:   0
  Abstain:  0
  Approval: 100.0%
  Threshold: simple_majority

Result: APPROVED

Completed in ~30s

Three voter roles deliberate via whichever local CLIs you have (Claude, Codex, Gemini) — no API keys required. Per-voter reasoning is recorded; the terminal prints the verdict. Mixed outcomes (some approve / some reject) and graceful error handling are demonstrated on the project site hero with a real 7-voter run.

4. Wire into your editor

nexus-agents setup   # Auto-configures MCP server in Claude Code, Cursor, etc.

Restart your editor. The 46 MCP tools (orchestrate, consensus_vote, research_synthesize, verify_audit_chain, …) become available to whatever agent you're already using.

What `setup` configures

By default, setup writes/updates up to seven things in your environment. Each can be skipped with the corresponding --skip-* flag if you don't want it.

Configured	Where written	Opt-out flag
MCP server registration (Claude)	`~/.claude/mcp.json` / Claude Desktop config	`--skip-mcp`
Project rules	`.cursor/rules/` and/or `.claude/rules/`	`--skip-rules`
Session hooks	`~/.claude/hooks/` (session-start / pre-tool / etc.)	`--skip-hooks`
OpenCode MCP config	`~/.config/opencode/opencode.json`	`--skip-opencode`
Gemini MCP config	`~/.gemini/mcp.json`	`--skip-gemini`
Codex MCP config	`~/.codex/config.toml`	`--skip-codex`
Project config file	`./nexus-agents.yaml`	`--skip-config`

Run with --interactive (the default) for a per-step confirm flow, or --no-interactive to accept all defaults.

5. Standalone usage (no editor required)

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 (focused synthetic dataset, n=10): 100% bug-catch, 50% raw FP rate; manual triage reclassified one of two inspected FP cases as a real finding (directional small-n, not measured rates) (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. Tamper-evident hash-chained append-only storage (tamper-evident, not tamper-proof — see threat model); integrity verifiable via `verify_audit_chain` MCP tool
Closed-Loop Telemetry	`OutcomeStore` feeds LinUCB + TOPSIS scoring; a second bounded, audited self-tuning loop demotes unhealthy CLIs (capped, auto-decaying, on by default, opt-out `NEXUS_TUNE_ENFORCE=false`)
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
46 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. Start with run — the default entry point: give it a goal and the MetaOrchestrator picks (and, with execute: true, runs) the right strategy. The other pipeline tools are advanced force-strategy paths for pinning a specific one.

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 per-node checkpoints + audit trail (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
`get_job_result`	Read result of an async-mode dispatch by jobId (#3042 / #2631)
`list_jobs`	List async-mode jobs across all tools — cross-session discovery (#3046 / #2631)
`cancel_job`	Mark an async-mode job as cancelled — idempotent (#3042 Stage 1b)
`ci_health_check`	CI infrastructure health — composes GitHub status + recent-runs activity (#3076)
`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
`run_quality_gate`	Run the QA quality gate (typecheck/lint/tests/build/security) over a project dir; returns structured pass/fail verdict + feedback
`suggest_research_tasks`	SUGGEST-ONLY: candidate pipeline tasks from research_discover findings for review — files/executes nothing (#1715)
`list_available_models`	Probe all model-discovery transports (OpenRouter API + opencode/claude/codex/gemini CLIs) and report per-transport health — validates the CLIs/APIs are reachable (#3406)
`run`	Default entry point — give a goal, MetaOrchestrator picks the strategy; returns the routing decision (execute:false, read-only) or runs it inline (execute:true; dev-pipeline+pipeline+research+consensus wired) (#3548)

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/nexus-substrate/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