Oficina global

English · Español

A Paperclip plugin that draws the agents of every company in your instance in a pixel-art office: who is working, who needs you, who failed, who is paused and who is idle. You can answer their questions without leaving the office, get spoken notifications and leave it running on a TV.

The interface and the voices are in Spanish (Chilean voices included). This README quotes the on-screen labels in Spanish, with a translation the first time they appear.

• Plugin key: jstm.oficina-pixel · npm package: paperclip-plugin-oficina-global · version 0.9.7
• Tested with Paperclip 2026.824.1.

What it does

• One floor, one room per company. Each agent has a desk and a character. While working, it types with its issue ID in a speech bubble; when it needs you, it raises its hand and lines up in front of your desk; on an error it turns red; when paused it sleeps; when idle it goes to the break room.
• It moves. Agents walk between their desk and the break room. An issue handed from one agent to another travels in an envelope or is handed over at the table. Idle agents chat about what is actually going on in the office, and pets in the hallway react when you click them.
• Answer without leaving. The questions, confirmations, suggested tasks and approvals that agents leave for you («Te toca», your turn) are answered in a window inside the office, with your session. The office only writes when you send an answer.
• Notifications and voices (off until you turn them on). When an agent starts needing you, a bell rings and its voice says what it needs («te necesito», I need you; «tengo un error», I have an error; «terminé», I'm done; «empecé», I started), with a different voice for each agent.
• Choosing the voices (0.11.0). In «Voces» (Voices), «Quién habla» (Who speaks) sets the voice of the Resumen (the spoken summary), the voice of «Leer contexto» (reads an issue aloud) and, on its own page, the voice of each agent, with a ▶ to hear each one first. The two that read on the spot only accept voices that answer in a few seconds (Lorenzo, Catalina, Alex, Dora, Santa) or the browser's voice; the four clones can be chosen for agents only. Changing an agent's voice records its phrases in the background and switches when they are all ready, so it never goes silent in between.
• Room sign. Under each room, a sign tells what each issue is about, one at a time.
• Summary. The «Resumen» (summary) button reads out loud what got done, what is waiting for you, what is being worked on and what is queued.
• Read context. In the top bar of each issue, a «Leer contexto» (read context) button reads out loud what the issue is about and what you are being asked.
• 3D view (0.12.0). On a computer, the «Vista 3D» (3D view) button lets you walk through the same office in first person, starting at your desk: look around with the mouse and walk with W A S D. Each room has its walls, doors to the hallway and a ceiling with lamps; monitors can be read up close, an arrow points to whoever needs you and clicking someone opens their card, as in 2D. Agents' chats and handoffs show over their heads, and the light follows the viewer's clock: the sky through the windows, the sun reaching the floor and, at night, a dimmer office.
• TV mode. Full screen at 1920×1080, plus a link to open it on a TV with a session.
• Phone. The whole floor and, when you tap a room, that room zoomed in.

It shows up in the sidebar («Oficina global»), on its own page, as a Dashboard widget and in the top bar (on phones; on issue pages, on computers too).

Requirements

• Paperclip with plugins (tested with paperclipai 2026.824.1; the SDK is pinned to that version).
• To install from npm: npm on the machine where Paperclip runs, which uses it to download the package.
• To install from source: Node 22 and npm.
• Recorded voices: Supertonic ships with the plugin (it downloads its model once); the other engines need uv, ffmpeg and internet access (see Voices). Without audio, the office uses the browser's voice.
• Optional, to let Chrome's built-in AI write part of the chats and the summary: desktop Chrome with Gemini Nano (see Chrome AI).

Installation

Once, before installing: the CLI needs a user who owns the instance (installing requires board access).

npx [email protected] auth login --instance-admin --api-base http://<your-server>:3100

If you run it from a Paperclip agent's session, drop the agent's variables (env -u PAPERCLIP_API_KEY -u PAPERCLIP_RUN_ID npx paperclipai …): installing needs your session, not the agent's.

From npm

npx paperclipai plugin install paperclip-plugin-oficina-global --api-base http://<your-server>:3100

Paperclip downloads it to ~/.paperclip/plugins/node_modules/paperclip-plugin-oficina-global on the machine where it runs.

From source

Paperclip loads the plugin from a local folder and serves the interface from its dist/ui/. It is best to install it from a separate copy from the one you edit.

git clone https://github.com/jstm-co/oficina-global.git
cd oficina-global
npm ci
npm run build
npx paperclipai plugin install "$PWD" --local --api-base http://<your-server>:3100

Upgrading

If you installed from npm, upgrade downloads the latest published version:

npx paperclipai plugin upgrade jstm.oficina-pixel --api-base http://<your-server>:3100

If you installed from source:

git pull && npm ci && npm run build
npx paperclipai plugin upgrade jstm.oficina-pixel --api-base http://<your-server>:3100

Uninstalling

npx paperclipai plugin disable jstm.oficina-pixel     # hides it without uninstalling
npx paperclipai plugin uninstall jstm.oficina-pixel   # removes it; it stores no data in Paperclip

The plugin only asks for interface capabilities (ui.page.register, ui.sidebar.register, ui.dashboardWidget.register and ui.action.register). The interface reads everything with the user's session, like Paperclip's Inbox.

Voices

Voices are optional. The audio files are recorded once, on the computer where Paperclip runs, and the page only plays them. Anything without audio is spoken by the browser's voice.

In the office, open the «Voces» (voices) panel, pick the engines and press Preparar voces (prepare voices). The panel shows the progress: first it downloads whatever is missing (once), then it records the phrases of the agents you can see. From then on the voices stay up to date: when a new agent shows up or you rename one, the page tells the plugin and the new phrases are recorded on their own, with no command to run.

| Voice | Engine | What it takes | License | |---|---|---|---| | 10 voices (5 female, 5 male) | Supertonic 3 (int8, with sherpa-onnx-node) | Ships with the plugin: downloads 145 MB once and uses ~0.5 GB of memory while recording. Mac, Linux and Windows | OpenRAIL-M (the model); Apache 2.0 (sherpa-onnx) | | Lorenzo and Catalina (Chilean) | Microsoft Edge TTS (edge-tts) | uv, ffmpeg and internet access. Unofficial free service: Microsoft receives the text of every phrase (agent and issue names) | LGPL-3.0 (the library) | | Alex, Dora and Santa | Kokoro-82M (kokoro-onnx) | uv and ffmpeg; downloads 354 MB once | Apache 2.0 | | 2 Chilean clones | Chatterbox Multilingual (chatterbox-tts) | uv and ffmpeg; 4–5 GB of memory and ~10 min per agent | MIT | | 2 Chilean clones | Qwen3-TTS 0.6B (qwen-tts) | Same as Chatterbox | Apache 2.0 |

• Only the engines you pick are used. Edge is not used unless you pick it, so no text goes to Microsoft.
• Supertonic needs neither uv nor ffmpeg. The others run with uv run --with <package> and need uv and ffmpeg on the PATH Paperclip starts with; the panel says which one can't be used and why.
• The models go to ~/.cache/oficina-voces/ and are checked against their SHA-256 when downloaded.
• The audio files live outside the plugin, so an upgrade doesn't delete them: in ~/Library/Application Support/oficina-global/voces/ on a Mac, ~/.local/share/oficina-global/voces/ on Linux and %LOCALAPPDATA%\oficina-global\voces\ on Windows, in one folder per plugin install. On start, the plugin links (or copies) them into dist/ui/voces/ to serve them. The first time, it adopts the ones that were already there.
• Supertonic records WAV (about 7 times the size of an MP3, ~230 KB for a 5 s phrase); the others, MP3.
• The worker also builds on the fly the phrases that say the issue's name (in vivo/, deleted after 7 days), with the engine of each agent's voice.
• The audio files are served next to the plugin's interface without a session: they say the name of each agent and its company. Keep that in mind if your Paperclip is open to the internet.

You can also record them from the terminal, in the plugin's folder (your clone, or ~/.paperclip/plugins/node_modules/paperclip-plugin-oficina-global if you installed from npm). It uses the engines picked in the panel, or the ones in --motores:

npm run voces -- --api http://<your-server>:3100                        # reads the agents with the CLI session (~/.paperclip/auth.json), GET only
npm run voces -- --api http://<your-server>:3100 --motores supertonic    # supertonic, kokoro, edge, qwen3 or chatterbox, comma-separated
npm run voces -- --api http://<your-server>:3100 --solo-plan            # no recording: which voice each agent gets and how many phrases are missing

Chrome AI

In desktop Chrome, Gemini Nano (free and local) can write the ending of the break-room chats and parts of the summary. You pick it in the «Conversaciones» (chats) menu. A check throws out anything that changes the meaning or brings in words that do not exist; without the AI, the usual phrases are used.

Chrome's AI only exists on secure pages: if you open Paperclip over http://, add its address in chrome://flags/#unsafely-treat-insecure-origin-as-secure and relaunch Chrome.

Outside Chrome, the plugin uses the page's LanguageModel if it has one (the same API), and the menus call it «IA del computador» (the computer's AI). For example, Paperclip Dispatch, the Mac app for Paperclip, serves it with Apple Intelligence.

Adapting it to your instance

This needs an install from source.

• src/instancia.ts: the fixed character of some agents (by ID) and how proper names are pronounced by the voices. In this copy it is empty: each agent gets a character from its name, title and role.
• src/plano.ts: how many columns each room has and which companies get their own color, by prefix. It comes tuned for the rooms of our instance (prefixes BET, ELM, JST and MIG); the others are drawn with the default sizes and in gray.

Development

npm ci
npm test            # tests (vitest)
npm run typecheck
npm run build       # dist/: worker, manifest, interface, tv.html and the voice generator

The code is in src/ (the logic, without React or the DOM, kept apart from src/ui/), the tests in tests/ and the voice generator in generador/. Code comments are in Spanish and mention issues from our internal board («JST-123»), which are not public.

Known limits

• The Paperclip API routes it reads are not part of the SDK (they are all in src/fuentes.ts): they may change with a new Paperclip version.
• The operating system notification needs https://; over http://, Paperclip shows the notice inside the page.
• Sound and voices start with your first click on the tab after opening it.
• iPhone has no full screen for web pages: TV mode explains how to add it to the home screen. It was not tested in Safari or on a real iPhone.
• The 3D view needs WebGL 2 and is for computers only (not on phones or in TV mode).
• Edge TTS is an unofficial service: if Microsoft changes it, the Lorenzo and Catalina voices stop generating until edge-tts is updated.

Contributing and releases

Pull requests are welcome: see CONTRIBUTING.md. Every pull request runs the tests, the type check and the build on GitHub Actions. When a maintainer merges package changes, GitHub Actions calculates a UTC date-based version (YYYY.MDD.N), publishes it to npm with provenance, and creates its GitHub Release. Nobody bumps a version by hand or publishes from their own computer.

License and attributions

The code is by JSTM, under the MIT license. The sprites and the notification sound are original (built by the scripts in src/ui/assets/fuente/).

There are two exceptions, under CC BY-SA 4.0:

• generador/referencias/ref_hombre_cl.wav and ref_mujer_cl.wav: phrases from the Google Crowdsourced Chilean Spanish set (OpenSLR 71, on Hugging Face as ylacombe/google-chilean-spanish), © Google. The voice clones come from them.
• src/ia/palabrasFiltro.ts: a filter with the most frequent Spanish words in the OpenSubtitles 2018 subtitles, from FrequencyWords (Hermit Dave).

See AVISOS.md (in Spanish). This project is not affiliated with Paperclip.