Run AI coding agents on your server.
Control them from your phone, browser, or Telegram.
Manages persistent terminal sessions for AI coding agents. Controls them from a mobile app, web browser, or Telegram.
Sessions run on your server as PTY processes. Close the app, leave for hours — when you come back, the full history replays instantly via a 4 MB ring buffer. Claude sessions resume exactly where they left off.
Not just notifications — a true two-way terminal over Telegram. When Claude goes idle, it reads the structured JSONL output and renders multi-select prompts as inline keyboards. Reply to approve, reject, or type a response.
An open manifest.json spec for wrapping any CLI tool. Declare the command, config schema with typed fields, capabilities — OpenDray generates the UI form, maps config to CLI flags and env vars automatically.
Register Ollama, Groq, LM Studio, or any OpenAI-compatible endpoint. When spawning an OpenCode session, pick a provider — OpenDray generates a per-session config and injects it at launch time.
Register MCP servers once in the UI. At session spawn, OpenDray renders per-session temp config files and injects them via CLI args. No global config pollution. Temp files cleaned on session exit.
Go backend with the Flutter web build embedded via go:embed. One binary + PostgreSQL. All config lives in the database — editable from the app. Your code and credentials never leave your infrastructure.
Each agent is a plugin with a manifest.json declaring its CLI command, available models, and capabilities. The UI adapts: resume buttons for Claude, model pickers for OpenCode, approval-mode selectors for Codex.
Choose from Claude Code, Codex, Gemini, OpenCode, Qwen, or a plain shell. Select a project directory and model. If using Claude, pick which OAuth account. For OpenCode, pick an LLM provider endpoint. Hit Create & Start.
OpenDray spawns the CLI in a pseudo-terminal, injects MCP server configs and LLM provider credentials as temp files, and starts capturing output in a ring buffer. Idle detection fires after 8 seconds of silence — triggering Telegram notifications.
Reopen the app and the WebSocket reconnects, replaying full session history. Or stay away and let the Telegram bridge handle it: idle notifications arrive with the latest output, structured prompts render as inline keyboards, your replies go straight to the agent.
Every provider in OpenDray — agents, panels, tools — is defined by a
manifest.json. The manifest declares the CLI command, available models,
and a configSchema that drives the settings form in the app.
Config → CLI args + env vars — each config field has a
type (string, secret, select, number, boolean, args) and optional
cliFlag or envVar. When a session launches, OpenDray
reads the user's config and builds the final command line and environment automatically.
Boolean fields append flags. Select fields append flag + value. Secrets inject as env vars.
Conditional fields — dependsOn and
dependsVal let fields appear only when relevant. For example,
an API key field only shows when the auth type is set to "custom" instead of "env" or "oauth".
Two plugin types: cli spawns a PTY process and manages its lifecycle.
panel is a stateless HTTP API backed by config (file browser, database viewer, log tailer).
Both are stored in PostgreSQL, toggleable at runtime, and health-checked in the background.
{
"name": "my-agent",
"displayName": "My Custom Agent",
"type": "cli",
"icon": "🤖",
"cli": {
"command": "my-agent-cli",
"defaultArgs": ["--no-color"],
"detectCmd": "which my-agent-cli"
},
"capabilities": {
"models": [
{ "id": "fast", "name": "Fast Mode" },
{ "id": "deep", "name": "Deep Analysis" }
],
"supportsResume": false,
"supportsStream": true,
"supportsMcp": true,
"supportsImages": false,
"dynamicModels": false
},
"configSchema": [
{
"key": "apiKey",
"label": "API Key",
"type": "secret",
"envVar": "MY_AGENT_API_KEY",
"required": true,
"group": "auth"
},
{
"key": "model",
"label": "Model",
"type": "select",
"options": ["fast", "deep"],
"cliFlag": "--model",
"cliValue": true,
"default": "fast"
},
{
"key": "verbose",
"label": "Verbose output",
"type": "boolean",
"cliFlag": "--verbose"
}
]
}When a session goes idle (agent waiting for input), OpenDray reads the structured output — for Claude, it parses the JSONL response log directly — and sends the latest content to linked Telegram chats. Reply to the notification and your text goes straight to the agent's stdin.
Structured prompts — if Claude shows a permission dialog or multi-choice question, the bridge parses the prompt type and renders it as inline Telegram keyboards. Tap a button to approve, or check multiple boxes and hit Submit. The response is formatted exactly as the CLI expects it.
Event-driven, not polling — the bridge hooks into the session idle detector. No busy loops. Notifications fire within seconds of the agent going quiet. Output is diffed against the last sent snapshot so you only see genuinely new content.
Works on any device with Telegram. Set up with a BotFather token and an allowed chat ID list. Multiple chats can link to the same session.
Read
/sessions — list running sessions
/screen [1] — current screen snapshot
/tail 1 [n] — last N lines of output
/whoami — show your chat ID
Control
/link 1 — bind this chat to session #1
/unlink — remove the binding
/links — list all active bindings
/send 1 <text> — one-shot send without /link
/stop 1 — stop a running session
Quick keys (sent to linked session)
/cc /cd /tab /enter /yes /noGo backend embeds the Flutter web build via go:embed. PostgreSQL stores sessions, plugin configs, Claude accounts, and LLM provider endpoints. Agents are spawned as real PTY subprocesses with injected credentials and MCP configs.
$ git clone https://github.com/opendray/opendray.git
$ cd opendray && cp .env.example .env
$ make devStop SSH-ing into tmux sessions. Start piloting your AI agents from anywhere.