◆ Nexotao Orce

Claude Code, in your browser — plus a real multi-agent orchestration system. Self-host it on a VM and drive it from any device: chat with a single agent, or give a goal and watch a planner break it into a task DAG that agent teams execute in parallel.

If Claude Code is one agent at your terminal, Orce is a room full of them — planned, tracked, and budgeted from a clean web console.

Why

You love Claude Code but you're tied to a terminal on a machine that has to stay on. Nexotao Orce puts the same engine (the official Claude Agent SDK) behind a web UI you host yourself, so you can prompt from your phone, laptop, or tablet — and it adds an orchestration layer for work that's too big for a single agent.

Features

• 💬 Chat — full Claude Code: bash, edit, read, web, subagents. Streaming, with persistent sessions you can resume from any device.
• 🧠 Orce (orchestration) — give a goal; a planner agent decomposes it into a task DAG, assigns each task to the best agent, and a scheduler runs them in parallel where dependencies allow.
• 🎫 Task system — every task is a ticket (TASK-0001), with a global Tasks board, status, dependencies, cost, and a detail page (status reasoning · connections · timeline · output).
• ✅ Plan approval — review the proposed DAG and Approve & run (or discard) before anything executes, or leave auto-run on (the default) to skip the gate.
• 🤖 Custom agents — CRUD agents with their own role, system prompt, model, allowed tools, and optional isolated workspace for safe parallel writes.
• 💰 Budget hard-stop — set a USD cap per run; warns at 80%, halts new tasks at 100%.
• 📜 Immutable audit log — every run keeps a timeline of events.
• 📂 Multi-project — an onboarding wizard to start fresh (empty workspace + one agent) or import an existing project (with a graph build + AI-proposed agents tailored to your stack). Switch between projects; each keeps its own sessions, agents, and runs.
• 📁 Workspace explorer — browse and preview the files/projects your agents create.
• 🕸️ Knowledge-graph aware — agents get a graphify tool that maps the project's code (calls, imports, god nodes), so they orient via the graph instead of grepping blindly. Auto-built structurally (no LLM), refreshed each run. Optional — enable with uv tool install graphifyy --with mcp.
• 📊 Bento dashboard — sessions, cost, storage, recent activity at a glance.
• 🔀 Two AI providers — pick at onboarding (or in Settings): Nexotao (recommended — paste a key, choose a model) or Claude Code (local), which connects straight through the machine's own logged-in Claude Code (subscription, API key, or Bedrock/Vertex/Foundry — nothing to paste).
• ♻️ Runs in the background — orchestration runs and chat turns keep executing even if you refresh or close the tab; reopen and they reconnect where they left off.
• 🔌 Plug & play — an embedded local Postgres (PGlite) persists everything with zero setup. Point at a hosted Neon/Postgres anytime.
• 🔐 Single-operator auth — you create the password on first run (nothing printed to the terminal), 🐳 one-container Docker, and a one-command CLI.

| | | |---|---| | | | | | | | | |

Quick start

You need Node.js 20+. Model access is set up in the app — pick Nexotao (a key) or Claude Code (local) (your machine's own login). See Choosing a provider.

Option 1 — install the nexotao command (recommended)

npm install -g nexotao-orce
nexotao

Typing nexotao starts the server and opens the UI in your browser. On first run you create your own password in the app (nothing is printed to the terminal), then onboarding walks you through your first project — pick a folder (type a path or Browse… it), choose a provider, and go.

• Stop it: press Ctrl+C in the terminal.
• Forgot your password? run nexotao reset-password, then set a new one on the next launch.
• Custom port: nexotao --port 9000 · don't auto-open: nexotao --no-open
• No global install? Run it once with npx nexotao-orce.

Option 2 — Docker

git clone https://github.com/rizkyriyadi/nexotao-orchestrator.git && cd nexotao-orchestrator
cp .env.example .env       # set APP_PASSWORD, ANTHROPIC_API_KEY, (optional) DATABASE_URL
docker compose up -d --build
# → http://localhost:8787

Option 3 — from source / development (hot reload)

npm install
npm run dev:server   # API on :8787  (reads server/.env)
npm run dev:web      # UI on :5173, proxies /api → :8787

Choosing a provider

On first run — and anytime under Settings — you pick where model access comes from:

• Nexotao (recommended) — a hosted, Anthropic-compatible gateway. Paste your sk-nexo-… key and choose a model (Claude Opus/Sonnet, GPT, Grok, DeepSeek). Grab a key at nexotao.com · docs.
• Claude Code (local) — connects directly through the Claude Code already installed and logged in on this machine. Nothing to paste and no model to pick — it uses whatever your local claude uses (subscription login, ANTHROPIC_API_KEY, or Bedrock/Vertex/Foundry). If claude works in your terminal, this works. The app auto-detects it and shows Detected. Not installed yet? Install Claude Code and run claude to log in.

You can switch providers at any time; the choice is global to your install.

Configuration

Everything is optional — sensible defaults let it run out of the box (embedded storage; you set the password + provider in the app).

| Variable | What it does | |---|---| | APP_PASSWORD | Web login password. If unset, you create one in the app on first run. Reset with nexotao reset-password. | | Provider | Chosen in the app (Nexotao or local Claude Code) — no env var needed. | | ANTHROPIC_API_KEY | Optional. Used by the Claude Code (local) provider if that's how your machine's claude is set up. | | DATABASE_URL | Optional. Empty → an embedded local Postgres (PGlite) persists to ~/.nexotao-orce/db (zero setup). Set it to use a hosted Neon/Postgres. NEXOTAO_DB=memory forces ephemeral. | | AGENT_CWD | The agent's working directory (its "project home"). Defaults to a workspace folder. | | PERMISSION_MODE | acceptEdits (default) · default · plan · bypassPermissions (requires the dangerous-mode acknowledgement below). | | ORCE_PLAN_MAX_NODES | Maximum planner DAG size. Default 32; invalid output is retried once and then fails visibly. | | ORCE_PLAN_MAX_DEPTH | Maximum dependency-path depth. Default 8. | | PORT | Default 8787. | | NEXOTAO_DATA | Where the CLI stores secrets + workspace (default ~/.nexotao-orce). |

Free persistent storage: create a project at neon.tech and paste its connection string into DATABASE_URL.

How it works

• server/ — Node + Hono. Wraps @anthropic-ai/claude-agent-sdk, normalizes its stream into small SSE events, and adds the Orce engine (planner → DAG scheduler), auth, and a Neon/Postgres (or in-memory) store. Serves the built UI, so it's a single process on one port.
• web/ — React + Vite + Tailwind. Streaming chat, the orchestration board, task pages, agents, files, and dashboard.
• Orce engine — planPhase validates task shape, exact agent names, unique keys, dependencies, graph size/depth, and acyclicity before persistence. Invalid model output gets one structured correction retry and then fails visibly. executePhase revalidates persisted plans, runs ready tasks in deterministic order, and terminates any no-ready-task state instead of waiting silently.

⚠️ Security

This server can run shell commands and edit files on its host. Treat it like SSH:

• Always put it behind HTTPS (e.g. Caddy: reverse_proxy localhost:8787). Never expose the raw port.
• Use a strong password (you set it on first run). It's single-operator by design — don't share it.
• It defaults to acceptEdits. To use unrestricted auto mode, set both PERMISSION_MODE=bypassPermissions and DANGEROUS_MODE_ACKNOWLEDGEMENT=I_UNDERSTAND_ORCE_CAN_EXECUTE_ARBITRARY_COMMANDS. Lock the VM firewall to your IP either way.

Roadmap

Reliability comes before feature breadth. Read the public roadmap, core lifecycle invariants, failure matrix, and measured baseline. Remaining restart, isolation, and hard-budget limitations are stated there explicitly.

Local diagnostics

Orce sends no telemetry. An authenticated operator may explicitly download an aggregate local report from /api/diagnostics/export. It covers install, onboarding, first plan, first successful run, seven-day return usage, and run outcomes without prompts, paths, source content, or secrets. See the data contract.

Contributing

Issues and PRs welcome — see CONTRIBUTING.md, SECURITY.md, and CODE_OF_CONDUCT.md.

License

MIT © Rizky Riyadi.

Built on the Claude Agent SDK. Not affiliated with Anthropic.