---
title: AgentBridge
emoji: 🤖
colorFrom: blue
colorTo: green
sdk: static
pinned: true
license: other
short_description: Self-hosted AI office agents — terminal chat, OpenAI-compatible API and MCP in one process
tags:
  - agent
  - agents
  - office-automation
  - rag
  - self-hosted
  - llm
  - tts
  - mcp
  - telegram
  - sip
---

# AgentBridge (codename "AGENT")

**Add agentic power to your work.** AgentBridge is a self-hosted server that runs AI agents with
two interfaces in a single process: a full-screen chat terminal (TUI) and a standard HTTP API
compatible with OpenAI — plus a native MCP connector. Chat in the terminal while scripts, bots and
apps use the same agents on the same port: same process, same conversations, no bridges, no
synchronization.

It encapsulates agentic functions in a well-designed **application-level sandbox** that prevents an
AI from acting outside its tools, while letting it handle real business tasks with a very high level
of privacy and confidentiality. Your documents stay on your machine: the knowledge base is a plain
folder on disk, not a cloud upload.

## What it can do

- **Documents, spreadsheets, presentations, PDF reports** generated end-to-end from a plain prompt
  (Word/Excel/PowerPoint via the OfficeTool plugin family).
- **Email** — read and send mail with your account.
- **Web research** — search, sources and reports.
- **Voice** — dictate and hear the assistant speak with in-process **Kokoro neural TTS** (no setup,
  voices included).
- **Phone access (SIP)** — the server becomes a phone endpoint, reachable 24/7 behind a DTMF PIN.
- **Telegram** — private-chat messaging with the agents, attachments both ways.
- **Scheduled tasks** — the agent plans work ahead and runs recurring duties on its own.
- **Podcasts** — complete podcast episodes (research → script → narration → MP3 + RSS) from one request.
- **Deterministic Memory** — a proprietary retrieval that keeps the agent grounded in everything it
  has already worked on, without burning reasoning iterations on recollection.

## Key technologies

| Technology | What it means |
|---|---|
| **AIOrchestrator library model** | One runtime, one agent core, one operational flow across terminal and API — no multi-server MCP overhead. Read the [white paper](docs/AIORCHESTRATOR-WHITEPAPER.md). |
| **Universal Tool System (UTS)** | Tools are compiled .NET assemblies the agent drives natively — no Python/Node, no interpreters, in-process with zero overhead, hot-add plugins. |
| **True application-level sandbox** | A structural action perimeter no tool method can breach — no chroot, Docker or VMs required. |
| **GDPR-ready anonymization** | Personal data is masked before any request reaches an external LLM provider and restored in the reply. |
| **Kokoro neural TTS** | Talks the moment you run it — in the terminal, over the API and on the phone. |

## Demo gallery

- [Terminal UI demo](media/demo.mp4) — streaming replies, `/` command palette, status bar
- [Presentation generated by the agent](media/demo1.gif)
- [PDF market & financial analysis report](media/enanced_doc_demo.mp4)
- [Invoice generated end-to-end](media/invoice.png) · [Employment contract](media/contract.png)
- [Excel spreadsheet with chart, single A4 page](media/spreadsheet-demo.png)
- [Podcast sample episode (MP4, ~42 s)](media/podcast-example.mp4) · [MP3](media/podcast-example.mp3)
- [Telegram agent chat](media/telegram-agent.png)

## Documentation

- **[User manual](docs/MANUAL.md)** — start here: install, configure the JSON files, use the TUI, connect OpenAI/MCP clients
- **[Terminal UI reference](docs/TUI.md)** — every command, shortcut and mouse action
- **[HTTP API reference](docs/API.md)** — chat, sessions, LLM switching, TTS, voice, files, MCP connector
- **[Beginner guides](docs/INDEX.md)** — 15 short plain-language guides (getting started → podcasts)
- **[Telegram chat](docs/telegram.md)** · **[SIP telephony](docs/sip.md)** · **[Text-to-speech](docs/TTS.md)** · **[Auto-update](docs/autoupdate.md)**
- **[AIOrchestrator white paper](docs/AIORCHESTRATOR-WHITEPAPER.md)** — why the library model outperforms traditional MCP deployments

## Install

Self-contained archives (~460 MB, no .NET needed) run on Windows x64, Linux x64/ARM64 and
macOS (Intel / Apple Silicon), and work with local models (Ollama, ExLlamaV2) or cloud providers
(DeepSeek, Z.ai, Gemini, Anthropic).

```bash
# Windows (PowerShell)
irm https://graphenelab.it/AgentBridge/install.ps1 | iex

# Linux / macOS
curl -fsSL https://graphenelab.it/AgentBridge/install.sh | bash
```

Or grab the archive for your platform from the
[Releases page](https://github.com/Graphene-Lab/AgentBridge/releases/latest) and run
`agent.exe` / `agent`.

## Source & license

- **Source code:** [github.com/Graphene-Lab/AgentBridge](https://github.com/Graphene-Lab/AgentBridge)
- **License:** [Andrea Bruno License 1.4](https://github.com/Graphene-Lab/AgentBridge/blob/master/LICENSE.md)
- **Website:** [graphenelab.it/AgentBridge](https://graphenelab.it/AgentBridge/)