Telegram plugin MCP connection drops after each conversation turn
Resolved 💬 3 comments Opened May 25, 2026 by Deepwater1000 Closed Jun 9, 2026
Bug Report
Description
When running claude --channels plugin:telegram@claude-plugins-official, the Telegram MCP plugin connection (plugin:telegram:telegram) repeatedly disconnects after processing a conversation turn. This causes the bot to become unresponsive to new messages until the session is restarted.
Environment
- Claude Code v2.1.150
- macOS Darwin 25.3.0
- Plugin:
telegram@claude-plugins-officialv0.0.6 - Started via:
tmux new-session -d -s claude-tg "claude --channels plugin:telegram@claude-plugins-official"
Reproduction Steps
- Start Claude Code with Telegram channel:
claude --channels plugin:telegram@claude-plugins-official - Send a message to the bot via Telegram
- Bot receives and processes the message successfully
- After Claude Code finishes its turn and returns to idle prompt (
❯), check/mcpstatus plugin:telegram:telegramshows✘ failed- Subsequent Telegram messages are not received by the Claude Code session
Observed Behavior
- The MCP connection is
✔ connectedimmediately after startup - After processing 1-2 messages (especially if tool use is involved like WebSearch), the connection drops to
✘ failed - The bun polling process (
bun server.ts) continues running but messages are not routed to the Claude session - Restarting the tmux session restores connectivity temporarily
MCP Status Before/After
# After startup (working):
plugin:telegram:telegram · ✔ connected · 4 tools
# After processing a message (broken):
plugin:telegram:telegram · ✘ failed
Other MCP servers that also fail in the same session
composio · ✘ failedplugin:github:github · ✘ failed
Workaround
Kill and restart the entire tmux session after each disconnection. This is not sustainable for a continuously running bot.
Frequency
Happens consistently — reproduced ~6 times in a single session (2025-05-25). The connection typically drops within 1-2 minutes of processing a message.
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗