Skip to main content
Glama
finished -a Claude "Login page fixed and tests pass"

🔔 Claude: Finished — Login page fixed and tests pass  ·  🔊 an anime girl (Shikoku Metan, sexy style) reads it to you


Why

You start an agent, switch to something else, and keep checking whether it is done. With finished, the agent tells you as soon as it stops. You see what it finished in the notification and hear it in one sentence, so you can stay away from the screen.

Related MCP server: mcp-windows-notify

Features

  • One call does both. It shows a Windows notification and speaks the message. The call returns in about 0.15 s and the speech plays in the background, so the agent is never held up.

  • Real anime voices. 18 VOICEVOX characters (Shikoku Metan, Kyushu Sora, Zundamon, ...) in sexy, whisper and sweet styles, running locally. They read English with a Japanese accent, like in anime.

  • 241 voices in all. Plus 49 local Kokoro voices (whisper, sultry, deep male, British, ...), 172 online neural voices in 75 languages, and Windows' built-in voices as a last resort.

  • Local first. The default voices run offline on your CPU with Kokoro-82M, and nothing leaves your computer.

  • Language detection. The language is detected offline with Lingua. A Serbian message is read by a Serbian voice, a German one by a German voice. Agents can also name the language themselves, which is exact.

  • Ready phrase, the agent's own words, or both. "Finished, senpai!", "Login page fixed", or "Finished. Login page fixed.".

  • No talking over each other. When several agents finish at the same time, they speak one after another.

  • Fast repeats. Everything spoken is cached, so a phrase you hear often plays at once.

  • Fallbacks. If the internet is down, a local voice speaks. If that fails, the Windows voice speaks.

  • A tray app. An English settings window to pick the voice, phrase and languages, listen to voices and see which agents are connected. While finished is on, its icon sits in the tray. Closing the window keeps it running; Quit in the tray turns it off.

  • Works automatically. The installer connects every agent installed on your computer with three things: the MCP tool, an end-of-turn hook (Claude Code, Codex), and one rule in its instructions. You hear about every finished task, even when an agent forgets to call the tool, and never twice. finished unregister undoes it all.

Install

You need Windows 10 or 11, Python 3.10–3.13 or uv, and about 400 MB of disk space.

git clone https://github.com/spixyy7/agent-finished
cd agent-finished
powershell -ExecutionPolicy Bypass -File install.ps1

Connecting your agents is required. finished only works when your agents call it, so the installer always connects them (see Connect your agents). It first says what it will change and asks once. If you answer no, nothing is installed. Every changed file is backed up first, and finished uninstall undoes it all. Pass -Yes to agree up front, for example in scripts.

The installer:

  1. creates .venv and installs the Python packages

  2. downloads the Kokoro model (about 350 MB, once) into models/

  3. puts a finished command in %USERPROFILE%\.local\bin for PowerShell, cmd and Git Bash, and adds that folder to your PATH

  4. connects every agent installed on your computer: the MCP tool, automatic notifications, and one rule

  5. asks whether to add the anime voices (VOICEVOX, about 1.9 GB; -NoAnime skips them, finished voicevox install adds them later)

  6. adds finished to the Start menu, starts it with Windows, and starts it now in the tray

When it is done, you hear "finished is installed" and the icon is in the tray. Restart your agents, then open finished from the Start menu to choose your voice.

The app

Open finished from the Start menu (or run finished app). Every change is saved right away.

  • Tray. While finished is on, its icon is in the tray. Click it to open the window. Right-click it for Send test notification, Speak on/off and Quit finished.

  • Closing the window hides it. finished stays on in the tray.

  • Quit finished turns it off. Agents' calls and hooks stay silent until you open the app again.

  • Start with Windows puts it back in the tray when you sign in. It does not turn finished on again after a Quit.

From a terminal: finished app (open), finished on (start in the tray), finished quit (turn off).

Quick start

finished app                         # the settings window
finished setup                       # the same choices as a text menu in the terminal
finished "Database migration done"   # what an agent runs
finished preview anime               # hear all anime voices (VOICEVOX)
finished voices                      # list voices

Connect your agents

The installer does this for you. To run it again or connect one more agent:

finished register            # every agent found on this computer
finished register codex      # just one
finished unregister          # remove everything it added (all agents, or name one)

For each agent it adds up to three things:

  • Tool: the MCP tool notify_finished, so the agent can say what it finished in its own words.

  • Auto: a hook that fires when the agent's turn ends, so you hear something even when the agent forgets to call the tool.

  • Rule: one short rule in the agent's global instructions, telling it to call the tool with a short summary when it is done.

Agent

Tool

Auto

Rule

Claude Code

claude mcp add -s user

Stop hook in ~/.claude/settings.json

~/.claude/CLAUDE.md

Codex CLI

~/.codex/config.toml

notify in ~/.codex/config.toml

~/.codex/AGENTS.md

Gemini CLI / Antigravity

~/.gemini/settings.json

~/.gemini/GEMINI.md

Qwen Code

~/.qwen/settings.json

~/.qwen/QWEN.md

OpenCode

~/.config/opencode/opencode.json

~/.config/opencode/AGENTS.md

Windsurf

~/.codeium/windsurf/mcp_config.json

~/.codeium/windsurf/memories/global_rules.md

Cursor

~/.cursor/mcp.json

Grok CLI

~/.grok/config.toml

  • Only agents installed on your computer are touched, and running register twice adds nothing twice.

  • Each file is backed up next to itself (*.bak-<time>) before it is changed.

  • unregister puts text files back byte for byte. JSON files keep the same content but get standard indenting.

  • If Codex already has a notify program (Codex's own computer-use helper, for example), it keeps working: finished runs it with the same data, and unregister restores the exact line that was there.

  • Restart the agents afterwards so they load the changes.

Automatic notifications

When a turn ends, the hook reads the agent's last reply and speaks its first sentence as the message, for example "Fixed the login bug and all 42 tests pass.". If the agent already called notify_finished during that turn, the hook stays quiet, so you never hear the same thing twice. The hook returns in about 0.1 s and never blocks or changes the agent.

finished hook claude --dry < stop-input.json      # what it would say, without speaking

By hand

The server is <folder>\.venv\Scripts\python.exe <folder>\finished.py mcp, using stdio.

{
  "mcpServers": {
    "finished": {
      "command": "C:\\path\\to\\agent-finished\\.venv\\Scripts\\python.exe",
      "args": ["C:\\path\\to\\agent-finished\\finished.py", "mcp"]
    }
  }
}
[mcp_servers.finished]
command = 'C:\path\to\agent-finished\.venv\Scripts\python.exe'
args = ['C:\path\to\agent-finished\finished.py', "mcp"]

Any agent that can run a shell command can use the CLI:

finished -a Aider -l en "Refactor done, 42 tests green"

The tool

notify_finished(message, language?, agent?)

Field

Meaning

message

What the agent finished, at most 12 words, no code or paths. Example: "Login page fixed and tests pass"

language

ISO 639-1 code of the message, such as en, sr or de. Optional, but it makes the voice choice exact.

agent

Shown in the notification title, such as Claude: Finished.

The tool description and the rule tell agents to call it once, when the whole task is finished or when they need you. This is the rule register adds, between <!-- finished:start --> and <!-- finished:end --> markers so it can be removed cleanly:

## Spoken notification (finished)
When you finish a task the user gave you, or you are blocked and need the user, call the
`notify_finished` tool once with a very short summary (max 12 words, plain words, no code or
paths) and `language` set to the summary's ISO 639-1 code. If the tool is not available, run:
finished -a <your name> -l <code> "<summary>"

Voices

finished voices             # anime and English voices, grouped
finished voices sr          # voices of one language
finished voices all         # all of them
finished preview anime      # hear a group: anime | sexy | female | male | local | online | <code> | <name>
finished set voice metan-whisper

Anime voices (offline, VOICEVOX, finished voicevox install):

Character

Voices

Shikoku Metan

metan-sexy (default), metan-whisper, metan-hush, metan-sweet, metan-tsun

Kyushu Sora

sora-sexy, sora-whisper, sora-sweet

Zundamon

zundamon-sexy, zundamon-whisper, zundamon-sweet

Others

tsumugi, himari, sayo, nurse-secret (a whispered secret)

Men

ryusei, takehiro, kotaro

These are VOICEVOX's own character voices with their セクシー (sexy), ささやき (whisper) and あまあま (sweet) styles. Japanese text is read natively and English words are read with a Japanese accent ("Login page fixed" becomes ロギン ペエジ フィックスト). The engine starts by itself the first time it is needed (about 280 MB of memory) and stops when you quit finished. Try the Japanese ready phrases too: おわったよ♡ ("I'm done"), お疲れさま♡ ("good work"), できたよ、ご主人様♡ ("it's done, master").

Local voices (offline, Kokoro):

Group

Voices

Sultry, whisper

whisper, sultry, bella, heart

Japanese accent

jp-soft, jp-whisper, jp-girl, jp-cute, jp-fox, jp-gentle, jp-bella, jp-heart

Other women

sky, nova, sarah, jessica, aoede, kore, river, emma-uk, isabella-uk, lily-uk, alice-uk

Men

deep, jp-boy, adam, michael, fenrir, puck, onyx, echo, eric, liam, santa, george-uk, fable-uk, lewis-uk, daniel-uk

Other languages

es-dora-local, es-alex-local, fr-siwis-local, it-sara-local, it-nicola-local, pt-dora-local, pt-alex-local, hi-alpha-local, hi-beta-local, hi-omega-local, hi-psi-local

The jp-* voices are Kokoro's Japanese voices speaking English, some blended with the whispering af_nicole voice. Without VOICEVOX, the default voice is heart.

Online voices need internet and come from Microsoft Edge's neural voices: ava, ava-sexy, emma, jenny, aria, sonia-uk, andrew, brian, christopher, guy, ryan-uk, plus 2–6 native voices for each of 75 languages (sr-sophie, de-tanja, ja-nanami, ...).

Windows voices are zira and david. They are robotic but always there.

Languages and auto-detect

finished languages              # all 75 languages, which ones auto-detect covers, which voice speaks each
finished set detect en,sr       # auto-detect chooses only between these (most precise)
finished set detect all         # or among all 60 detectable languages
finished set language de        # or always one language
finished set voice-for sr sr-nicholas   # which voice reads Serbian
finished detect "Sredio sam bug u checkoutu"   # see what a text is detected as

The language is chosen in this order:

  1. The agent names it (language in MCP, -l in the CLI). This is used as is and is always right.

  2. A fixed language set with finished set language <code>.

  3. Offline detection with Lingua, among the languages you allowed with set detect:

    • Fewer allowed languages means higher accuracy. Choosing between English and Serbian is much easier than choosing among 60.

    • Languages that short texts cannot tell apart count as one group: Serbian / Croatian / Bosnian, and Norwegian Bokmål / Nynorsk. So Latin-script Serbian is not read as Croatian.

    • A single word keeps your main voice's language unless the detector is sure. Two- or three-letter words like OK or API need 95% certainty.

    • In both mode, the phrase and the message are detected separately, so "Finished" + "Završio sam login" uses two voices.

Read aloud: only your main voice's language. Some languages only have older online voices that sound robotic. With finished set speak main (Read aloud → Only English in the app), only messages in your main voice's language are read. A message in another language makes the voice say your ready phrase, and the notification still shows the whole message. Agents are then told to write their summary in that language.

Then the voice is chosen in this order: your main voice if it speaks that language, then your pick for that language (voice-for), then a native online voice of the same gender as your main voice, then a local Kokoro voice of that language, then your main voice.

The included tests check 23 short English and Serbian messages (including single words) and 21 sentences in 21 languages with all languages enabled. All pass.

Settings

Command

What it does

finished set voice NAME

Main voice. It reads every message in its own language.

finished set voice-for CODE NAME

Voice for another language (default resets it).

finished set language auto|CODE

Detect each message, or always use one language.

finished set detect all|en,sr,...

Languages auto-detect chooses from.

finished set speak all|main

Read every language in its own voice, or only your main voice's language (others say the ready phrase).

finished set mode preset|agent|both

Say the ready phrase, the agent's message, or both.

finished set phrase N|"text"

Pick a ready phrase (finished phrases) or write your own.

finished set volume 0-200

Volume in percent.

finished set toast on|off

Turn the notification on or off.

finished set sound on|off

Turn the voice on or off.

finished set words 3-40

Most words spoken from a message (default 14).

finished config

Show the settings (stored in config.json).

finished app / on / quit

Open the window, start in the tray, or turn finished off.

CLI options for a single call: -a AGENT, -l CODE, -v VOICE, -p N|"phrase", -w (wait until spoken), --no-voice, --no-toast.

How it works

agent ──► notify_finished (MCP) ──┐
agent ──► finished "msg"  (shell) ┼─► background worker ─┬─► notification (WinRT toast via Windows PowerShell)
turn ends ─► hook (Claude, Codex) ┘                      └─► language → voice → speech ─► cache ─► speaker
             (skipped if the agent already called the tool)        Kokoro (local) · Edge (online) · SAPI (Windows)
  • finished.py contains the CLI, voices, language detection, speech and the MCP server.

  • agents.py contains register / unregister and the end-of-turn hooks.

  • app.py contains the settings window and the tray icon (Tk with the Sun Valley theme, pystray).

  • toast.ps1 shows the notification. It registers its own app name, finished, so you can mute it in Windows settings.

  • languages.json lists the languages and online voices. Rebuild it with python tools/build_languages.py.

  • cache/ holds the spoken WAV files (the newest 500) and finished.log.

Privacy

  • With local voices (VOICEVOX, Kokoro, Windows), nothing leaves your computer.

  • With online voices, the message text is sent to Microsoft's Edge read-aloud service through edge-tts. Messages in languages without a local voice use an online voice. To stay fully offline, run finished set language en (or keep detect to languages that have local voices) and pick a local main voice.

  • There is no telemetry and no account.

Troubleshooting

Problem

Try

No sound

Run finished -w "test" and read cache/finished.log.

Nothing at all

finished is off (you quit it). Open it from the Start menu, or run finished on.

No notification

Turn off Focus / Do Not Disturb, and check Settings → System → Notifications → finished.

A language sounds robotic

Use Read aloud → Only English (finished set speak main), or pick another voice for it (finished voices CODE).

The first message takes about 3 s

That is the model loading. Phrases are cached, and later messages are faster.

A message is read in the wrong language

Narrow it down with finished set detect en,xx, or have the agent pass language.

The agent does not see the tool

Run finished register again and restart the agent.

You hear the reply's first line but want the agent's own summary

The rule asks agents to call the tool. Check that it is in the agent's instructions (finished register).

Start over

Delete config.json and cache/.

Uninstall

finished uninstall           # quit, take finished out of every agent, remove the shortcut, autostart and command

Then delete the folder. Backups of every agent file it changed are kept next to them (*.bak-<time>). To only disconnect agents, use finished unregister.

Development

.venv\Scripts\python.exe -m unittest discover tests     # detection, voices, speaking rules, MCP, register round trip, hooks
.venv\Scripts\python.exe tools\build_languages.py        # refresh the language and voice list

Credits

  • VOICEVOX by Hiroshiba Kazuyuki and contributors (engine LGPL-3.0), the anime voices. Each character has its own terms. They are free to use, and if you publish audio made with them, credit the character, for example VOICEVOX:四国めたん. finished does not ship VOICEVOX; finished voicevox install downloads the official engine.

  • Kokoro-82M by hexgrad (Apache-2.0), the local voices

  • kokoro-onnx by thewh1teagle (MIT), which runs Kokoro with ONNX Runtime

  • edge-tts by rany2 (LGPL-3.0), the online voices

  • Lingua by Peter M. Stahl (Apache-2.0), language detection

  • python-soundfile (BSD-3-Clause)

  • pystray (LGPL-3.0), Pillow (MIT-CMU) and Sun Valley ttk theme (MIT), the app

This project is not affiliated with Microsoft, Anthropic, OpenAI, Google or any agent vendor.

License

MIT

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      D
      quality
      D
      maintenance
      Enables MCP agents to send desktop notifications for completed tasks, failures, and important updates, providing visible and reliable feedback.
      1
      54 npm
      2
      Apache 2.0
    • A
      license
      B
      quality
      D
      maintenance
      Enables AI development tools on Windows to send native system notifications for task completion, errors, and reminders, with customizable icons and sounds.
      7
      20 npm
      8
      MIT
    • F
      license
      A
      quality
      D
      maintenance
      Provides a notify_done tool for Claude Desktop to trigger desktop popup notifications when tasks complete.
      1
      -