v2.155.0 MIT-licensed Node 22+ Works with Claude · Codex · Gemini · OpenCode

A control plane for AI coding agents.

Your agents already write the code — they're the data plane. Nexus Agents is the control plane above them: it admits work through one entry point, reviews it adversarially before it ships, records every action in a tamper-evident hash-chained audit log, and routes the next task based on what actually worked. One MCP server. Every CLI you already have.

Run the 60-second proof Source on GitHub Installation

~/work · nexus-agents vote · 2026-05-17

$ nexus-agents vote --proposal "Add a 'nexus-agents tour' interactive
    walkthrough command that runs after 'setup'..."

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

Proposal: Add a 'nexus-agents tour' interactive walkthrough command that runs after 'setup'...

Collecting votes from 7 agents (timeout: 300s each)...

Votes

  ✓ Software Architect:   APPROVE (82%)
  ✓ Security Engineer:    APPROVE (90%)
  ✓ Developer Experience: APPROVE (88%)
  ✗ AI/ML Engineer:       REJECT  (75%)
  ✓ Product Manager:      APPROVE (82%)
  ✗ Contrarian Analyst:   REJECT  (95%)
  ✗ Scope Steward:        ERROR   — MCP error -32001: Request timed out

Summary

  Approve:  4
  Reject:   2
  Abstain:  0
  Errored:  1
  Approval: 66.7%
  Threshold: simple_majority

Result: APPROVED

Completed in 188487ms

Verbatim from a real run on 2026-05-17 (a proposal from the website backlog; the panel said yes 4 to 2). Seven specialised voter roles deliberate via your local Claude, Codex, and Gemini CLIs — no API keys required to start. Pass --record <issue-number> to write the result back to GitHub as a structured comment; pass --verbose to print per-vote verification hashes.

The 60-second proof.

Three commands. No keys to set up just to look. The hard part is later.

1 Install

npm install -g nexus-agents
nexus-agents doctor

expected A health table; rows that need API keys are marked missing. doctor only reads; it never writes.

2 Run a real consensus vote (no API keys needed)

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

expected Output in the same shape as the hero transcript above, but with 3 voters instead of 7 (--quick). Roles deliberate via your local Claude/Codex/Gemini CLIs (whichever you have installed). Approval, reject, abstain counts + final decision print to the terminal.

3 Attach to your editor

# Auto-configure as an MCP server for Claude Code, Cursor, etc.
nexus-agents setup

expected The MCP server registers in your editor's config. Restart the editor, and 46 new tools (orchestrate, consensus_vote, research_synthesize, verify_audit_chain, …) become available to whatever agent you're already using.

That's the smoke task. Every command above is configurable, but the defaults already work — nexus-agents orchestrate "…" will pick the right CLI, run the right experts, and write the trace.

What's in the box.

admission control

Adversarial review before it ships.

The gate that catches what a data-plane agent would otherwise ship. pr_review runs voter roles against a verification gate; consensus_vote offers 5 aggregation strategies — simple majority, supermajority, unanimous, proof-of-learning, higher-order Bayesian — across seven specialised roles. Every decision is recorded with per-voter reasoning.

event log

Tamper-evident hash-chained audit.

Every model call, every vote, every routing choice flows through an append-only, hash-chained audit log on disk. Replay any decision; integrity is verifiable via verify_audit_chain. It is tamper-evident, not tamper-proof — alteration is detectable, not impossible. Drift detection fires when conventions slip.

closed loop

Route the next task on evidence.

A 12-stage pipeline (budget → capability → preference → TOPSIS → LinUCB → latency) picks the right CLI from the 5 adapters you have installed. Outcomes feed 8 memory backends so tomorrow's routing learns from today's failures. A second bounded loop demotes unhealthy CLIs automatically — capped, auto-decaying, demotion-only. Higher authority is earned against evidence, never flipped on by default.

Plus 11 expert personae, 9 research sources, and 176+ tracked papers feeding the planner. The full inventory lives in the architecture pages.

Architecture, briefly.

Control plane and data plane. Work is admitted through MCP, scheduled to whichever CLI scored best, reviewed by consensus, and recorded in the audit log — then outcomes feed back so the control plane routes the next task on what actually worked.

flowchart LR mcp["MCP Protocol 46 tools"] orch["Orchestrator task analysis"] router["Composite Router · 12 stages Budget → Capability → Preference TOPSIS → LinUCB → Latency"] clis["Claude · Gemini · Codex · OpenCode"] outcome[("Outcome Store feeds back")] subgraph services ["Supporting"] consensus["Consensus 5 strategies"] memory["Memory 8 backends"] audit["Audit chain"] end mcp --> orch --> router --> clis --> outcome outcome -.->|learns routing| router orch -.-> services

Next steps.

Action items first; reference last. Pick any one — they don't have to be in order.

try

A control plane for AI coding agents.

The 60-second proof.

What's in the box.

Adversarial review before it ships.

Tamper-evident hash-chained audit.

Route the next task on evidence.

Architecture, briefly.

Next steps.

Run an orchestration.

Connect Claude Code, Cursor, or any MCP client.

Run the full dev pipeline.

Tune routing & experts.

Architecture deep dive.

Browse the research registry.