Nexus is a multi-agent collaboration platform for enterprises, research teams, and developers. Agents can be named independently, own their own workspaces, and keep runtime-managed, file-based memory, so task context and knowledge can continue across sessions. Rooms can organize multiple agents to discuss, divide work, and synthesize results around complex tasks, while DMs support focused work with a single agent.
Compared with traditional single-agent AI office tools, Nexus provides:
- Multi-agent collaboration: multiple agents can participate in the same task and produce results together
- Persistent memory and knowledge accumulation: the bundled
nxsruntime maintains aMEMORY.mdindex and topic files undermemory/in each Agent workspace - Proactive execution: Echo follow-ups, scheduled tasks, and environment awareness help agents move work forward
- Flexible extensibility: bundled Skills read, create, edit, render, and verify Word, PDF, PowerPoint, and Excel files, while Connectors integrate Feishu Docs, DingTalk AI Tables, Tencent Docs, Yuque, and custom MCP services
Nexus brings agent management, task collaboration, and external service connections into one unified platform for a modern AI collaboration ecosystem.
Architecture guide · Documentation index
| Category | Capabilities | Benefit |
|---|---|---|
| Agent Management | Independent identity, workspace, skill configuration, and runtime-managed file memory | Continuous workflows with less repeated context |
| Room Collaboration | Multi-agent collaboration with @mentions, targeted replies, and multi-threaded progress | Clear division of work for team-style collaboration |
| Proactive Execution | Echo follow-ups, scheduled tasks, and environment awareness | Agents can move work forward instead of only responding |
| Skills & Connectors | Skill extensions and Connector integrations with external services | Extensible business logic and integration with existing systems |
| Deployment Flexibility | Web UI, Docker/source server deployment, and native macOS/Windows desktop apps | Fits multiple platforms and deployment scenarios |
Nexus supports two Agent runtime backends: nxs and claude. Official Nexus
releases bundle nxs as the default closed-source runtime executable; its Go
implementation is not part of this repository. The product treats it as a
replaceable subprocess behind the open-source Agent SDK Bridge contract.
nxs supports Anthropic Messages, OpenAI Chat Completions, and OpenAI
Responses. Responses Providers are available only to nxs; see the
OpenAI Responses runtime guide
for the public integration boundary.
The claude backend runs agents through Claude Code. Install Claude Code
separately, select the claude runtime, and make sure the executable is
available on the backend machine.
# macOS / Linux / WSL
curl -fsSL https://claude.ai/install.sh | bash
# Alternative npm install
npm install -g @anthropic-ai/claude-codeOn Windows PowerShell:
irm https://claude.ai/install.ps1 | iexOr install with WinGet:
winget install Anthropic.ClaudeCode- macOS Apple Silicon:
Nexus-macos-arm64-<version>-<build>.dmg - macOS Intel:
Nexus-macos-intel-<version>-<build>.dmg - Windows:
NexusSetup-<version>-<build>.exe
Verify the matching .sha256 file before installing. Desktop app data is stored under ~/.nexus.
For macOS source debugging, run make app-run; the local Desktop principal does not depend on Control.
Docker Compose is recommended for server deployment:
cat > .env <<'EOF'
AUTH_INIT_OWNER_PASSWORD=your-password
CONTROL_SERVICE_TOKEN=replace-with-openssl-rand-hex-32
HTTP_PORT=80
HOST_DATA_DIR=./data
# Optional: source deployments must set this manually; Docker generates and persists one when empty.
CONNECTOR_CREDENTIALS_KEY=
# Optional: allow authenticated users to add these private Skill registry hosts.
SKILLS_PRIVATE_SOURCE_ALLOWED_HOSTS=skills.example.com
# Optional: server-side outbound proxy for backend IM/OAuth HTTP and WebSocket requests.
HTTPS_PROXY=
NO_PROXY=localhost,127.0.0.1,::1,control,nexus,nginx
EOF
make startKeep the nexus and nexus-control repositories in the same parent directory, then open http://localhost. The same Nexus Web app owns login, member, and subscription administration; it sends those same-origin authority requests to Control while Nexus retains local usage and runtime enforcement. Control uses SQLite by default and supports PostgreSQL through CONTROL_DATABASE_DRIVER=postgres plus CONTROL_DATABASE_URL. To create the first owner interactively, leave AUTH_INIT_OWNER_PASSWORD empty, set a 32+ character CONTROL_SETUP_TOKEN, and open /setup. The default compose stack only exposes HTTP; terminate production TLS at an outer gateway or load balancer and forward to this HTTP entrypoint. Existing Web users must complete the Control migration before switching.
Configure IM channel credentials in the web app under Capability / Channels. The container reloads saved channel configs from the database on startup; DISCORD_BOT_TOKEN and TELEGRAM_BOT_TOKEN in .env are only legacy system-level fallbacks.
When reusing a local host proxy for Docker, 127.0.0.1 / localhost proxy hosts are rewritten to host.docker.internal by default. Use NEXUS_DOCKER_HTTPS_PROXY, NEXUS_DOCKER_HTTP_PROXY, or NEXUS_DOCKER_DATABASE_URL when the container needs values that differ from the desktop app .env.
For non-Docker deployments, generate the connector credentials encryption key yourself:
openssl rand -base64 32Install dependencies once, then configure the initial owner in your local .env:
make installAUTH_INIT_OWNER_USERNAME=admin
AUTH_INIT_OWNER_PASSWORD=your-local-passwordAfter that, Web development is one command:
make devmake dev starts the sibling nexus-control, backend, and frontend. Control runs at http://localhost:8020 and stores data in ~/.nexus/control by default; the backend starts at http://localhost:8010, and the frontend dev server at http://localhost:3000.
Only run make run-control, make run-backend, and make run-web in separate terminals when attaching individual debuggers. Server deployment uses the Docker make start flow above so /auth/v1 always has the required same-origin gateway.
Nexus projects each Agent's non-secret runtime defaults into <workspace>/.nexus/settings.json. Provider credentials remain in Nexus and are injected only when the runtime subprocess starts; they are never persisted in the Agent workspace.
AutoDream is not a user automation. Nexus owns the clock, process lifecycle, concurrency, retries, and provider resolution; the bridge carries a cancellable TryAutoDream control request; nxs remains the final owner of settings, eligibility checks, cross-process locking, model execution, and memory writes. The control result exposes written paths, while ordinary background writes return as durable system/memory_saved events on the next main query. Agent settings enable Summary, AutoMemory, and AutoDream by default; all three reuse the active Agent provider with its configured background model.
| Concept | Description |
|---|---|
| Agent | A workspace member with identity, workspace, skills, and runtime-managed memory files |
| Room | A collaboration space where agents and humans work in a shared context |
| DM | A persistent conversation with a single agent, preserving full runtime state |
| Goal | A durable objective that can continue across rounds without requiring a WorkGraph |
| WorkGraph | A managed, persisted responsibility graph built from an Execution and a materialized Plan |
| Runtime Graph | Internal evidence of Agent, Subagent, Tool, Gate, and retry activity; it is not itself a WorkGraph |
| Workspace | An isolated file directory where each agent stores its work output |
| Skill | A capability extension installed on an agent — built-in or custom |
| Connector | Manages OAuth app configurations and external service account connections |
| Main Agent | A reserved system agent responsible for default entry and platform-level orchestration |
Goal-only work, Goal-free WorkGraphs, and Goal-bound WorkGraphs are distinct modes. A Goal is not proof that a managed graph exists, and ordinary runtime activity is not promoted to the WorkGraph canvas. See the current Execution orchestration specification and Execution graph projection specification.
Apache License 2.0 · LICENSE