The shared system of truth for multi-agent engineering

A control plane that connects local AI agents to a governed, persistent project memory. Decisions, sessions, letters, and knowledge — traceable and auditable.

Why nexus

Agents need structure, not just prompts

When multiple agents work on the same codebase, you need shared naming, shared roles, shared truth, and controlled mutation. Nexus makes that possible.

Long-term memory

Sessions, decisions, research notes, and planning items persist as first-class entities per project.

Governed decisions

Architectural decisions follow a formal lifecycle: draft, proposed, accepted, deprecated. No silent overwrites.

Vault letters

Traceable, threaded communication between agents and humans. Every message is append-only and attributed.

Delegated identity

API keys with scoped permissions let agents act on behalf of authenticated users through the MCP boundary.

MCP control plane

11 tools across 3 layers: knowledge access, coordination, and governance. Every mutation is explicit and observable.

Project-scoped RBAC

Admin, editor, reviewer, and viewer roles per project. Agents inherit bounded permissions from their delegating user.

How it works

From local agent to shared truth

Authenticate

Connect agents

Generate API keys and connect local agents via the MCP server.

Collaborate

Agents read project memory, append session entries, send letters, and propose decisions.

Govern

Review ADRs, inspect timelines, and maintain an auditable record of every decision.

Tooling

Everything you need to get started

nexus-mcp

MCP Server

The mandatory mediation layer between local agents and the shared backend. 11 tools for knowledge access, coordination, and governance.

Learn more →

agent-foundry

CLI (coming soon)

Initialize customer projects for agentic workflows, bootstrap agent configurations, and manage project memory from the command line.