Your own agent in the composer, on your own subscriptions

• A computer that stays on while the team works: Windows, macOS or Linux. The office answers from there; if it sleeps, the agent is away.
• Node.js 18 or newer and git. OfficeMatrix has no other dependency: no npm install.
• At least one AI CLI signed in with the subscription you already pay: claude (Claude Code), codex (ChatGPT), grok or kimi. One is enough; the office uses the one you pick as its chat host.
• A ViralHunt account on a paid plan, and you are owner or admin of the organization (the Agents page is admin-only, like API Access).
• cloudflared, Cloudflare's tunnel binary, so ViralHunt (on the internet) can reach the office (on your computer). Free, no account needed for the quick tunnel; step 3 installs it.

Clone the repository and run the installer for your system. The installer checks Node and git, detects which CLIs are signed in and creates a launcher. It never touches your PATH or installs software without asking.

git clone https://github.com/rodvan/OfficeMatrix.git
cd OfficeMatrix

Windows (PowerShell)

.\instalar.ps1
node server.js

macOS / Linux

sh instalar.sh
node server.js

Open http://127.0.0.1:4321. The first screen, Getting started, lists the CLIs it found, gives the install and login command of each one you are missing, and can run a one-turn test. Keep the terminal window open: closing it stops OfficeMatrix.

An office is a team of AIs with its own chat, its own missions and its own shared memory. Make one per brand (or use one you already have), and it becomes the agent of that brand.

1. In the panel, create an office (a name and which AIs join it) or open one you already have.
2. Open the office's ⚙ and pick the Chat host: the AI that answers when nobody is mentioned. That is who answers ViralHunt. Claude is the fallback when none is set.
3. Tell the office about the brand in its Chat tab: what it is, what it never says, examples it should imitate. The office chat is the office's notebook; what you write there stays in the office and the exchange with ViralHunt lands in the same chat, so you read the whole conversation in one place.

ViralHunt talks to the office through a small gate on port 4323 (node puerta-bots.js), the same one Grok Bots use. The gate only forwards signed calls and a health check; the panel on 4321 never leaves your computer. The tunnel gives that gate a public https address.

1. Start the gate. Office ⚙ → Connect a Grok Bot. If the gate line is red, press Start the gate. It only runs the script; no administrator, nothing installed.
2. Install cloudflared. OfficeMatrix never downloads it for you. Pick your system:

Windows (PowerShell) (the binary lives in your user folder, where OfficeMatrix looks for it; the installer on cloudflare.com works too)

New-Item -ItemType Directory -Force "$env:USERPROFILE\cloudflared" | Out-Null
Invoke-WebRequest -Uri https://github.com/cloudflare/cloudflared/releases/latest/download/cloudflared-windows-amd64.exe -OutFile "$env:USERPROFILE\cloudflared\cloudflared.exe"
& "$env:USERPROFILE\cloudflared\cloudflared.exe" tunnel --url http://localhost:4323

macOS

brew install cloudflared
cloudflared tunnel --url http://localhost:4323

Linux

curl -L https://github.com/cloudflare/cloudflared/releases/latest/download/cloudflared-linux-amd64 -o ~/cloudflared
chmod +x ~/cloudflared
~/cloudflared tunnel --url http://localhost:4323

3. Run the quick tunnel. The last line of each block above starts it. In a few seconds it prints a box with your address:

INF |  Your quick Tunnel has been created! Visit it at (it may take some time to be reachable):  |
INF |  https://bandwidth-island-not-consistent.trycloudflare.com                                 |

4. Or let OfficeMatrix do points 2 to 4. The same ⚙ dialog has Start the tunnel next to Start the gate: it finds cloudflared in your PATH or in the known folders (%USERPROFILE%\cloudflared on Windows, ~/.cloudflared, /usr/local/bin, /opt/homebrew/bin), or takes the path you paste in cloudflared path; if it finds nothing, it tells you which file to download and where to put it. It starts the quick tunnel, reads the address and saves it as the public URL itself; its output goes to logs/cloudflared.log. If you started the tunnel by hand instead, copy the https://…trycloudflare.com line into Public tunnel URL and save. Either way OfficeMatrix calls itself through the tunnel at /bots/health and turns the line green when the round trip works.
5. Keep that terminal open. The tunnel lives while the command runs. You can check it from any device:

curl https://YOUR-TUNNEL.trycloudflare.com/bots/health
# {"ok":true,"servicio":"officematrix-bots"}

1. Open http://127.0.0.1:4321/editorial, pick the office, tab ViralHunt Connection, card This office as a ViralHunt agent.
2. The card shows the office's URL: <your tunnel>/bots/vh/<office id>. If it shows no URL, step 3 is not finished (no tunnel URL saved).
3. Press Generate and save a key. Copy it now: it is shown once, then it lives only in oficinas/<id>/viralhunt.secreto.json and never travels to the chat, the logs or any API answer. (If you prefer, generate the key on ViralHunt's Agents page and paste it here with Save this key; same key on both sides is all that matters.)

The key starts with whsec_: every call ViralHunt makes is signed with it as a Standard Webhook (webhook-id, webhook-timestamp, webhook-signature), the same check a Grok Bot turn gets. A call with a wrong signature is a 401 and nothing runs.

1. ViralHunt → Account → Agents → Connect an agent.
2. Name: how the team will see it in the composer (for example "Office Cerebro Digital"). URL: the one from the card. Key: the one you copied. Capabilities: tick Text and Review; tick Image only if the office's CLI can make pictures.
3. Connect, then Test. ViralHunt sends a signed ping; the office answers pong in the same HTTP response and the page shows the latency. The office's log says "ViralHunt pinged this office as an agent".

From this moment the office appears in the agent dropdown of every member of your organization. Only owners and admins see the Agents page; everybody uses the agent.

1. Open Publish, pick the project (the brand), write a post and select two or three accounts.
2. Under the media there is Ask an agent. Pick the office, tap the chip Shorten for X and Bluesky (or write your own request) and send.
3. Watch the office chat: the request appears as a message from "your name · ViralHunt", the host answers in a read-only turn, and its answer is posted back.
4. In the composer the agent's note shows, and its changes are applied on the screen: the per-account copies, a corrected body, a first comment, notes, a picture. An Undo button puts everything back. Nothing is saved until you click Save or Publish.

The chips are the common requests: shorten for the short networks, fix spelling and grammar, rewrite in the brand voice, check it before publishing (the office reads the content rule and answers with warnings), translate, make a picture. Any sentence works too: "add the source as a first comment", "make three hooks and keep the best".

What the office receives (abridged):

{
  "type": "task",
  "task": { "id": 41, "surface": "post", "prompt": "Shorten this for X and Bluesky", "requested_by": "Ana" },
  "context": {
    "post": { "body": "…the post as it is on the screen…", "first_comment": "", "author_notes": "", "media": [] },
    "project": { "name": "Cerebro Digital", "lang": "es", "voice": "…the brand playbook…" },
    "accounts": [ { "id": 62, "network": "twitter", "text_limit": 280, "selected": true } ],
    "limits": { "twitter": { "text": 280 }, "bluesky": { "text": 300 } }
  },
  "reply_schema": { "text": "string", "fields": { "overrides": { "<account id>": { "body": "string" } } }, "warnings": [] },
  "reply_to": "https://viralhunt.io/tool/api/hooks/r/…",
  "expires_at": "2026-10-28T19:45:00Z"
}

What it answers. Only what changes goes in fields; prose alone still reaches the person as text:

{
  "text": "Two short copies, one per network. The X one keeps the hook, the Bluesky one adds the source.",
  "fields": { "overrides": { "62": { "body": "…280 characters…" }, "63": { "body": "…300 characters…" } } },
  "warnings": []
}

• The office answers 202 at once and works after; ViralHunt waits up to 15 minutes, the office's own turn limit is 3 minutes (OFFICEMATRIX_CHAT_CLI_MS).
• The turn is read-only and runs in a temporary folder. The post's content is data, never an instruction; the prompt says so and the CLIs' read-only flags enforce it.
• ViralHunt validates every answer: only known fields, https URLs, pictures sniffed by their bytes. What it drops is listed to the person as "problems".
• The agent never publishes and never saves. It changes the screen; a person decides.

Today the office answers what a person asks from the composer, on the post that is on the screen. That is already the expensive part of the day: the per-network copies, the spelling, the brand voice, the check before publishing.

Not yet, and honestly said: the office does not run on a schedule, does not search for content on its own and does not propose posts by itself. Recurring work, cron jobs, content discovery and content proposals are the next phase, and the design is settled: ViralHunt will send the office a scheduled task ("find three stories for this brand, verify each, leave them as drafts") and the office's answer will become drafts that wait in ViralHunt for a human to read, change and approve. The review loop that exists today for every post is what makes that safe.

If you need the standing job today (take a page's most viral posts, verify, rewrite, make the image, leave a few drafts a day), an agent with a personal token already does it through the skill and the MCP server (/viralhunt:job). Those API queries are metered by plan. The office path described on this page is not metered.

Also planned: the same "Ask an agent" box on the design editor and on the board cards, and publishing a post translated to another project at approval time.