opendray
Run Claude Code, Codex, Antigravity, Grok Build & OpenCode on your own server, and drive them from web, mobile, or chat.
🌐 English · 简体中文 · فارسی · Español · Português · 日本語 · 한국어 · Français · Deutsch · Русский
Running an AI coding agent over SSH means it dies the moment your laptop sleeps. opendray runs it on a host that stays awake (a Mac mini, a NAS, a VPS) and lets you reattach from a browser, a phone, or a chat message. The session keeps executing whether or not anyone is connected.
One Go binary. Your infrastructure. No cloud account, no telemetry, no subscription.
Why opendray#
- Sessions that never drop. The agent runs on your host and survives client disconnect, laptop sleep, and host reboot. Reattach hours later; the transcript is exactly where you left it.
- Drive from anywhere. A React web admin, a Flutter mobile app, and six chat channels (Telegram, Slack, Discord, Feishu, DingTalk, WeCom), every one bidirectional. Get pinged when a session goes idle; reply to feed the next prompt back in.
- Five agents, one gateway. Claude Code, Codex, Antigravity, Grok Build, OpenCode, plus any shell. Adding a CLI is a JSON descriptor drop-in.
- Multi-account fleet. Drop several logged-in accounts on the host; opendray pools them, balances new sessions, and switches a live session between accounts without losing the conversation.
- Local-first memory. Cross-session recall across user / project / session scopes, embeddings from ONNX, Ollama, or LM Studio. No vector data leaves your network.
- Built to build on. REST + WebSocket API with scoped keys, per-call audit log, and reverse-proxy mounts. Use opendray as your own product's agent runtime.
- Self-hosted & licence-clear. Apache 2.0, one static binary, cosign-signed releases.
Quick start#
Linux / macOS / WSL2. The installer walks through Postgres, CLI install, admin credentials, and service registration (~5–10 min):
[object Promise]Windows sets up WSL2, then runs the same installer inside it:
[object Promise]Just the binary (bring your own Postgres + supervisor):
[object Promise]Then open the web admin at http://<host>:8770/admin/.
📖 Prefer to do it by hand? The 15-minute getting-started guide mirrors each step. Binary-only setup lives in docs/install-binary.md; production deploy (systemd / launchd) in docs/getting-started.md.
Features#
| Sessions | Attach to a running Claude Code, Codex, Antigravity, Grok Build, OpenCode, or shell session from web, mobile, or chat. Survive client disconnect and host reboot. Image attachments, TUI theme-following, mouse-wheel scroll in full-screen TUIs. |
| Round Table (experimental) | Seat several providers plus yourself in one shared thread. @mention who replies (or @all); each answers in character after reading the whole conversation. Summarize it, or convert it into a role-assigned, multi-session execution plan. |
| Providers | 5 first-class AI coding CLIs plus arbitrary shell. New CLI = a JSON descriptor under internal/catalog/builtin/. Per-provider MCP-server injection. One-click version checks + updates. |
| Memory | Three-scope retrieval (user, project, session). Local-first embeddings via ONNX, Ollama, or LM Studio. Cross-layer conflict detection. Knowledge pages injected at spawn. |
| Database | Inspect and query project databases from the session inspector: PostgreSQL, MySQL, MariaDB, SQLite. Per-project isolation. Web and mobile. |
| Channels | Telegram, Slack, Discord, Feishu, DingTalk, WeCom, plus a Bridge adapter for custom transports. Bidirectional: sessions notify, replies feed back. |
| Integrations | REST + WebSocket API with scoped keys, per-call audit log, reverse-proxy mounts. HashiCorp Vault MCP. See integration guide. |
| Ops & security | Single Go binary. One-line installer (Linux, macOS, WSL2). Self-managing (opendray update / start / stop). Encrypted Postgres backups. Cosign-signed releases + SPDX SBOM. Apache 2.0, no telemetry. |
How it compares#
| opendray | Claude Desktop | Cursor | CLI over SSH | ChatGPT Desktop | |
|---|---|---|---|---|---|
| Session survives client disconnect | ✅ | ❌ | ❌ | ⚠️ tmux/screen | ❌ |
| Multi-account pool with live switch | ✅ | ❌ | ❌ | ❌ | ❌ |
| Cross-session memory layer | ✅ | ❌ | Partial | ❌ | Partial |
| Host filesystem + tool use | ✅ | Limited | ✅ | ✅ | Limited |
| Mobile client with feature parity | ✅ | ❌ | ❌ | ⚠️ SSH client | Partial |
| Chat channel adaptors | ✅ (6) | ❌ | ❌ | ❌ | ❌ |
| Self-hosted | ✅ | ❌ | ❌ | ✅ | ❌ |
| Licence | Apache 2.0 | Proprietary | Proprietary | (varies) | Proprietary |
Against self-hosted chat frontends (Open WebUI, LibreChat, Dify): opendray runs the actual agent CLI with tool use and file writes on the host, in a reattachable PTY, not just a chat box in front of an API.
Who it's for#
- Homelab solo dev. You have a box running 24/7 and you're tired of Claude Code dying when your laptop sleeps. Keep it running; reattach from your phone.
- Small-team lead. Pool 3–5 accounts, watch per-account usage, hand teammates scoped keys and a mobile app with no App Store submission.
- Integrator. You need to spawn agent sessions with tool use and don't want to reimplement session lifecycle, PTY handling, memory, or channel routing. Treat opendray as your agent runtime.
How it works#
A single Go binary on your host serves the web admin (/admin/) and a REST + WebSocket API (/api/v1/*). Clients drive sessions over HTTP/WS; the session manager spawns each AI CLI in its own PTY; the memory layer keeps shared state in Postgres with vector embeddings from your own provider. Everything runs on your network, with no inference outside your control.
Note: opendray shares process state (
~/.claude, ssh-agent, project files) with the CLIs it spawns, which is incompatible with production container isolation, so Docker is not a supported deploy path for v2.x. Run it on a host (bare metal, VM, or LXC).
Production deploy#
The one-line installer sets up a service for you. To wire it up yourself, each path gives you auto-restart on crash, persistent state, and secrets kept out of config.
systemd (Linux, recommended). Ships a hardened unit at deploy/systemd/opendray.service (ProtectSystem=strict, NoNewPrivileges, migrate-then-serve boot):
macOS (Mac mini / Studio as home server) uses a LaunchDaemon, and any other supervisor (OpenRC, FreeBSD rc.d) works with the plain binary. Full walkthroughs (secrets file, launchd plist, updates) live in docs/install-binary.md.
Docs & community#
License#
Apache 2.0. No telemetry, no cloud account, no subscription.