simple-notify-mcp

Model Context Protocol (MCP) server for Codex and Claude Code with text-to-speech (TTS) and Telegram notifications.

Tools

• simple_notify_status: always available; returns capabilities, missing config, and setup-web availability/running state.

• simple_notify_setup_web_start: available when --enable-setup-web is set; starts the local setup web UI on demand and returns the current tokenized URL.

• simple_notify_setup_web_stop: available when --enable-setup-web is set; stops the local setup web UI when no longer needed.

• tts_say: text-only input; async by default, uses configured provider (openai, fal-minimax, fal-elevenlabs) with macOS say fallback.

• telegram_notify: available when Telegram bot token + chat id are configured; supports parse_mode (plain, markdown, html) and returns hasUnreadIncoming from a non-advancing unread peek.

• telegram_send_photo: available when Telegram bot token + chat id are configured; sends local image files (jpg, jpeg, png, webp, gif, bmp) with optional caption and parse_mode.

• telegram_read_incoming: available when Telegram bot token + chat id are configured; reads incoming updates for configured chat.

• telegram_read_media: available when Telegram bot token + chat id are configured; reads image updates and can return MCP image content blocks.

Telegram formatting quick examples:

• telegram_notify({ "text": "**Build done**. [Diff](https://example.com)", "parse_mode": "markdown" })

• telegram_send_photo({ "filePath": "/tmp/plan.png", "caption": "<b>Plan snapshot</b>", "parse_mode": "html" })

• Markdown mode supports a safe subset: **bold**, *italic*, _italic_, ~~strike~~, `code`, [text](https://url), and # headings.

• HTML mode is validated and allows only Telegram-safe tags; links must be https:// or http://.

• If Markdown entity parsing fails on Telegram side, the server retries once with plain text for reliability.

• Telegram limits are enforced before send (message: 4096 chars, caption: 1024 chars).

Install (npx)

1) Recommended: agent-managed setup web

Use this if you want easy reconfiguration anytime without keeping an HTTP port open all the time.

codex mcp remove simple-notify
codex mcp add simple-notify -- \
  npx -y simple-notify-mcp@latest \
  --enable-setup-web

Ask your agent (Codex / Claude Code / another agent) to run simple_notify_status, call simple_notify_setup_web_start if needed, and then send you setupWeb.url.

2) Minimal runtime: no setup web

Use this if config is already done and you do not want the setup server running.

codex mcp remove simple-notify
codex mcp add simple-notify -- npx -y simple-notify-mcp@latest

If you need to change provider/keys later, switch back to mode 1.

Optional: pass API keys via env when adding:

codex mcp add simple-notify -- \
  --env OPENAI_API_KEY="$OPENAI_API_KEY" \
  --env FAL_KEY="$FAL_KEY" \
  -- npx -y simple-notify-mcp@latest \
  --enable-setup-web \
  --setup-port 21420

Optional legacy behavior:

• add --setup-web-autostart if you explicitly want the setup web server to bind during MCP startup

Install (Claude Code)

1) Recommended: agent-managed setup web

Use this if you want easy reconfiguration anytime without keeping an HTTP port open all the time.

claude mcp remove simple-notify
claude mcp add --transport stdio simple-notify -- \
  npx -y simple-notify-mcp@latest \
  --enable-setup-web

Ask Claude Code to run simple_notify_status, call simple_notify_setup_web_start if needed, and then share setupWeb.url.

2) Minimal runtime: no setup web

Use this if config is already done and you do not want the setup server running.

claude mcp remove simple-notify
claude mcp add --transport stdio simple-notify -- npx -y simple-notify-mcp@latest

If you need to change provider/keys later, switch back to mode 1.

Optional: pass API keys via env when adding:

claude mcp add --transport stdio \
  --env OPENAI_API_KEY="$OPENAI_API_KEY" \
  --env FAL_KEY="$FAL_KEY" \
  simple-notify -- \
  npx -y simple-notify-mcp@latest \
  --enable-setup-web

Optional legacy behavior:

• add --setup-web-autostart if you explicitly want the setup web server to bind during MCP startup

How To Use

Think of simple-notify-mcp as your "agent communication layer":

• voice message when work is done (tts_say)

• Telegram ping when work is done or while work is in progress (telegram_notify)

• optional Telegram inbox reads (telegram_read_incoming, telegram_read_media)

Typical flow

1. Start with setup web enabled (recommended mode above). This exposes setup-web start/stop tools, but does not open a local port yet.

2. Ask your agent for setup link:

• "Run simple_notify_status. If setup web is not running, call simple_notify_setup_web_start and send me setupWeb.url."

3. Open the link, set keys/provider/chat id, then click Save. In some clients, you may need to restart the agent process after adding keys so all tools become available.

4. When you are done configuring, ask your agent to call simple_notify_setup_web_stop.

5. After that, ask your agent to always:

• speak on completion

• send Telegram on completion

• send Telegram updates during long tasks

What to ask your agent (examples)

• Setup:

  • "Please configure simple-notify and give me the setup link."

• Completion behavior:

  • "When you finish a task, call TTS and Telegram notify."

• Long task behavior:

  • "If task is long, notify me on progress milestones and before escalation requests."

• Read incoming Telegram:

  • "If notify result says unread incoming, read it and continue."

TTS/Notification style tips (easy and practical)

You can tune:

• language (EN/RU/etc)

• tone (calm/energetic/formal/casual)

• emotion (neutral/cheerful/serious)

• slang level

• pacing

Examples:

• EN calm:

  • Task complete. Build passed. I left a short summary.

• EN upbeat:

  • Done. All checks are green.

• RU neutral:

  • Готово. Проверки прошли успешно.

• RU casual:

  • Запилил фичу, всё пашет, тесты зеленые.

Copy-Paste For AGENTS.md / CLAUDE.md

The easiest way to tune agent behavior is to add explicit tool-usage instructions to your agent config. You can copy-paste this block and adjust it as needed:

If user uses simple-notify-mcp:

1) Setup flow
- Call simple_notify_status.
- If setupWeb.enabled=true and setupWeb.running=false, call simple_notify_setup_web_start.
- If setupWeb.running=true, return setupWeb.url to user.
- When setup is finished or user asks to close it, call simple_notify_setup_web_stop.
- If setup web is disabled, tell user to run MCP with --enable-setup-web.

2) Completion flow
- On task completion, call tts_say with a short completion message.
- Then call telegram_notify with a short completion summary.
- If telegram_notify returns hasUnreadIncoming=true, optionally call telegram_read_incoming.

3) Long-task flow
- For long tasks, send milestone progress via telegram_notify.
- Send a notify before asking user for escalation/approval.
- Keep updates useful (no spam).

4) Safety
- Never include secrets/tokens in TTS or Telegram messages.

Configuration Schema

Config file path (default):

• $XDG_CONFIG_HOME/simple-notify-mcp/config.json

• or ~/.config/simple-notify-mcp/config.json

Env key precedence:

• OPENAI_API_KEY overrides keys.openai.apiKey

• FAL_KEY / FAL_API_KEY override keys.fal.apiKey

{
  "tts": {
    "provider": "openai",
    "params": {
      "openai": {
        "model": "gpt-4o-mini-tts",
        "voice": "alloy",
        "speed": 1,
        "responseFormat": "mp3",
        "instructions": "Speak calmly and clearly."
      },
      "falMinimax": {
        "voiceId": "Wise_Woman",
        "speed": 1,
        "vol": 1,
        "pitch": 0,
        "emotion": "neutral",
        "englishNormalization": false,
        "languageBoost": "auto",
        "outputFormat": "url",
        "audioFormat": "mp3",
        "audioSampleRate": 32000,
        "audioChannel": 1,
        "audioBitrate": 128000,
        "normalizationEnabled": true,
        "normalizationTargetLoudness": -18,
        "normalizationTargetRange": 8,
        "normalizationTargetPeak": -0.5,
        "voiceModifyPitch": 0,
        "voiceModifyIntensity": 0,
        "voiceModifyTimbre": 0,
        "pronunciationToneList": [
          "燕少飞/(yan4)(shao3)(fei1)"
        ]
      },
      "falElevenlabs": {
        "voice": "Rachel",
        "stability": 0.5,
        "similarityBoost": 0.75,
        "style": 0,
        "speed": 1,
        "timestamps": false,
        "languageCode": "en",
        "applyTextNormalization": "auto"
      }
    }
  },
  "telegram": {
    "chatId": "123456789"
  },
  "keys": {
    "openai": {
      "apiKey": "sk-..."
    },
    "fal": {
      "apiKey": "fal_..."
    },
    "telegram": {
      "botToken": "123:ABC"
    }
  },
  "misc": {
    "ttsAsyncByDefault": true
  }
}

MiniMax voices available in setup UI:

• Wise_Woman, Friendly_Person, Inspirational_girl, Deep_Voice_Man, Calm_Woman, Casual_Guy, Lively_Girl, Patient_Man, Young_Knight, Determined_Man, Lovely_Girl, Decent_Boy, Imposing_Manner, Elegant_Man, Abbess, Sweet_Girl_2, Exuberant_Girl

Setup Web UI Flags (disabled by default)

Flags:

• --enable-setup-web (default off)

• --setup-web-autostart (default off; optional legacy eager-start behavior)

• --setup-host (default 127.0.0.1; non-loopback values are clamped to 127.0.0.1)

• --setup-port (default 21420)

• --setup-token (optional; if omitted, generated per run)

Behavior:

• --enable-setup-web exposes the on-demand setup-web tools but does not bind a port by itself

• simple_notify_setup_web_start starts the local setup server only when the agent needs it

• simple_notify_setup_web_stop closes the local setup server when you are done

• --setup-web-autostart restores eager startup if you explicitly want the previous behavior

• local bind only

• if --setup-port is occupied, server uses the next free local port

• setup URL includes the current run token query parameter

• if --setup-token is omitted, each fresh start generates a new token

• use simple_notify_status to discover whether setup web is running, plus setupWeb.url and missingConfig

Tool Contracts

simple_notify_status

Input:

{}

simple_notify_setup_web_start

Input:

{}

Output notes:

• starts setup web only when --enable-setup-web is set

• returns current setup-web state, including setupWeb.url

• if already running, returns the existing URL without rebinding

simple_notify_setup_web_stop

Input:

{}

Output notes:

• safe to call repeatedly

• returns wasRunning=false when setup web was already stopped

tts_say

Input:

{ "text": "Job done" }

Output notes:

• default mode is async (misc.ttsAsyncByDefault=true), so tool returns immediately after queuing speech

• set misc.ttsAsyncByDefault=false in setup web Misc tab for blocking/sync behavior

telegram_notify

Input:

{ "text": "Job done" }

Output notes:

• returns { "accepted": true } by default

• adds hasUnreadIncoming: true only when unread messages are detected

• performs a non-advancing unread peek (limit=6) before returning

telegram_read_incoming

Input:

{
  "limit": 20,
  "timeoutSeconds": 0,
  "advanceCursor": true
}

Output notes:

• reads updates filtered to configured telegram.chatId

• tracks cursor in memory for the current server run

• set advanceCursor=false to peek without moving cursor

telegram_read_media

Input:

{
  "limit": 20,
  "timeoutSeconds": 0,
  "advanceCursor": true,
  "includeData": true,
  "maxImages": 1,
  "maxBytesPerImage": 8000000
}

Output notes:

• only image media is returned (text-only updates are ignored)

• when includeData=true, tool can return MCP image content blocks (base64 + mime type)

• large files are skipped based on maxBytesPerImage

• media cursor is tracked in memory for current server run

Self-test

npm run self-test -- --text "Task complete. Build passed and your results are ready."

Disable one side:

npm run self-test -- --no-tts
npm run self-test -- --no-telegram

Notes

• tts_say is text-only; provider/model/voice/etc are server config.

• tts_say runs async by default; switch in setup web Misc tab if you need sync mode.

• OpenAI and FAL network errors fall back to macOS say when available.

• OpenAI responseFormat=pcm is not directly playable by this local player path.

• Runtime config is reloaded from disk before tool calls, so manual config edits are picked up without restarting the MCP process.