Skip to content

Security

  • User whitelist — only approved Telegram IDs can interact
  • Admins & guests — a shared bot can reserve approvals, lifecycle and transport for ADMIN_USER_IDS; see Sharing a bot
  • Group confinement — RESTRICT_TO_GROUPS keeps guests out of private chats
  • Group roles — inside a group, admins / contributors / spectators, so being in the chat need not mean being able to prompt the agent (GROUP_MEMBERS_DEFAULT)
  • Addressed-only — GROUP_REQUIRE_MENTION keeps human conversation in a shared group from becoming prompts
  • Permission gate — dangerous Bash patterns pause for an admin’s approval, on the SDK and PTY transports alike (CCR included), subagents too
  • Scope guard — tool calls reaching outside the shared projects, or at any credential path, pause too
  • Charter judge — guest messages are read against a charter before they run
  • Project sandbox — the project picker is scoped to WORKSPACE_DIR
  • Permission mode — /mode sets how much Claude asks before acting
  • Dangerous mode — opt-in auto-approve for all tool permissions
  • Secrets — loaded from .env (gitignored), never committed
  • Secret requests — a key Claude asks for goes from your message into the command it named, and the message is deleted; Claude sees only the result. Private chats with an admin only
  • Gate fails closed — in PTY mode a tool runs only on the bot’s explicit allow: if the bot can’t be reached (restarting, erroring), the tool is blocked. Background agents and turns Claude starts on its own between your messages go through the gate too, in the chat the session belongs to
  • Local IPC token — the loopback server the hooks and MCP tools talk to requires a random per-process token, passed through the environment only, so other users on the machine can’t drive it

TeleCoder has no server of its own. Its state stays on your machine, in ~/.telecoder (owner-only permissions), and includes session metadata and preferences. Before November 2026 it was ~/.claudegram: the launcher moves that across on its first start and leaves a link at the old path. On a shared bot it also includes the Telegram ids and names of the people it has seen or let in.

Your messages go to Telegram and to Claude Code, and from there to the model provider you use. A secret you send when Claude asks for one is the exception: it goes to Telegram like any message, until the bot deletes it, and from there only into the command Claude named. A Claude token made with claude setup-token doesn’t go through Telegram at all. Claude Code keeps its own session transcripts in ~/.claude. Voice and text-to-speech go to Groq or OpenAI, and only if you configure them. Nothing goes to Telegraph unless you set TELEGRAPH_ENABLED=true; with it on, long replies, /handoff notes and action logs are published there, and those pages are public to anyone with the link.