pocket-opencode

A Telegram bot daemon that runs on your Windows PC and lets you drive the local opencode agent from your phone: send prompts, switch models/agents/projects, watch responses stream into a single Telegram message, approve tool permissions, and inspect diffs — all with zero open ports (Telegram long polling only).

How it works

Telegram (your phone)
   │  long polling
   ▼
pocket-opencode daemon (Bun + grammY)
   │  HTTP + SSE (localhost only)
   ▼
opencode serve --port 4096
   ▼
your projects

Requirements

• Bun ≥ 1.1
• opencode CLI on your PATH (opencode --version works)
• A Telegram bot token (create a bot via @BotFather)
• Your Telegram user ID (get it from @userinfobot)

Setup

1. bun install
2. Copy-Item config.example.json config.json (PowerShell) and fill in:
   • your bot token from @BotFather
   • your numeric Telegram user ID in "allow"
   • your project name/path in "projects"
3. bun run start

On first start the daemon prints a 6-digit pairing code in the terminal and the bot replies "Pairing required" to any message. Send /pair <code> in Telegram once — the bot is then bound to that chat. Only user IDs in "allow" are ever processed; everyone else is silently ignored.

Configuration

config.example.json is the template — copy it to config.json and replace the placeholders:

{
  "telegram_token": "123456:REPLACE_WITH_YOUR_TELEGRAM_BOT_TOKEN",
  "opencode_port": 4096,
  "db_path": "pocket.db",
  "allow": [111111111],
  "projects": [
    { "name": "myproject", "path": "C:/code/myproject" }
  ]
}

• telegram_token — the token @BotFather gave you when you created the bot.
• allow — array of Telegram user IDs allowed to talk to the bot. Your numeric ID comes from @userinfobot. This is the hard allowlist: nobody else is processed.
• projects — real folders on this PC; add more entries to switch between them with /cd.
• opencode_port / db_path — optional; defaults shown.

config.json is gitignored — your token never gets committed.

Commands

Command	Action
(plain text)	send a prompt to opencode in the active project
/status	active project, session, model/agent/thinking overrides, opencode health
/new	start a fresh opencode session
/model	pick a model (inline keyboard)
/agent	pick an agent (inline keyboard)
/think <low|medium|high|max>	store a thinking-level preference
/cd	switch project (inline keyboard; sessions kept per project)
/stop	abort the current run
/undo	revert opencode's last change
/diff	show changed files (+additions/-deletions)

Streaming

opencode emits message.part.updated SSE events; the daemon coalesces deltas (~1.2 s cadence) and edits one Telegram message in place — no message spam. session.idle finalizes the message; permission.updated surfaces an Allow/Deny inline keyboard wired to opencode's permission API.

Thinking levels

/think stores your preference and shows it in /status. The opencode SDK (1.18.x) does not expose a thinking/reasoning parameter on prompts, so it is applied best-effort wherever opencode supports it.

Security notes

• Hard allowlist: user IDs not in "allow" are dropped before any processing.
• First-run pairing code gates chat control (printed once to the terminal).
• opencode is talked to over 127.0.0.1 only; nothing listens on external interfaces.
• config.json (contains your bot token) is gitignored; *.db state files too.

Run

bun run start

Ctrl+C stops the daemon gracefully (bot stops, SQLite closes).