Agent Office

Agent Office is a self-hosted workspace where people and AI agents can collaborate on Word documents, spreadsheets, presentations, and PDFs. Upload or create a document, mention an agent in chat, and watch it work in the same file while progress streams back in real time.

• AI-assisted document editing — assign work with natural-language messages such as @Alice update the first paragraph.
• Real-time collaboration — chat, task progress, and document changes update live for workspace members.
• Office file support — work with Word, Excel, PowerPoint, OpenDocument, PDF, and text files.
• Flexible editor backend — run with Collabora Online or OnlyOffice.
• Workspace agents — configure different models, credentials, prompts, skills, and MCP servers per workspace.
• Self-hosted — keep accounts, documents, and workspace data in your own deployment.
• English and Traditional Chinese — switch languages from the sign-in page or sidebar.

1. Copy the environment file and update its credentials:

cp .env.example .env

2. Start Agent Office with your preferred editor:

# Collabora Online
docker compose -f docker-compose.yml -f docker-compose.collabora.yml up -d

# OnlyOffice
docker compose -f docker-compose.yml -f docker-compose.onlyoffice.yml up -d

3. Open http://localhost:8788, create an account, and create your first workspace.

Set a model credential in .env:

ANTHROPIC_API_KEY=
OPENAI_API_KEY=
GEMINI_API_KEY=

You only need the credential for the provider you plan to use. Claude subscription tokens are also supported through CLAUDE_CODE_OAUTH_TOKEN.

You can then create an agent from the workspace Agents page and choose its provider, model, prompt, skills, and credentials. Workspace settings take precedence over deployment defaults.

Open a document and assign work from its chat panel:

@Alice Change the date in the first paragraph to August 2026 and add a thank-you sentence at the end.

Typing @ opens a list of enabled agents. The selected agent opens the document in its own browser session, edits it, saves it, and reports the result in chat.

Start the backend:

cd backend/AgentOffice.API
dotnet run

Start the frontend in another terminal:

cd frontend
npm install
npm run dev

The frontend is available at http://localhost:5173, and the API runs at http://localhost:5000.

• Anthropic
• OpenAI and OpenAI Codex
• Google AI
• AWS Bedrock
• Google Vertex AI
• OpenRouter
• Other providers supported by the built-in pi runtime

• Architecture and security boundaries
• Known issues

• frontend/ — web interface
• backend/ — API, authentication, workspaces, documents, and WOPI host
• agent-worker/ — AI task runner and browser-based Office tools
• docs/ — architecture and operational notes

The interface supports English and Traditional Chinese. Language selection is remembered automatically. Translation files are located in frontend/src/locales/.