DeLive captures system audio from your desktop and routes it through a dozen different speech recognition backends — cloud and local — for real-time transcription. What makes it stand out is how it doesn’t stop at transcription alone; it exposes the transcripts and correction workflows through a local MCP server, turning your desktop audio into a programmable data source for AI agents. For anyone working with speech-to-text, agentic AI, or local-first workflows, DeLive offers a unique bridge between desktop audio and AI-driven automation.
What DeLive does and its architecture
At its core, DeLive is a cross-platform Electron desktop application written in TypeScript. It focuses on capturing system audio — that is, the raw audio output your machine is playing — and transcribing it in real time. Unlike many transcription tools that rely on a single ASR engine, DeLive integrates twelve Automatic Speech Recognition (ASR) backends. These range from cloud providers, including Deepgram and ElevenLabs, to local models like whisper.cpp.
This multi-backend approach is a key architectural choice. It lets users balance tradeoffs around cost, latency, privacy, and transcription quality by selecting or combining different ASR providers. For example, local whisper.cpp offers offline transcription without exposing audio data to the cloud but may have higher latency or lower accuracy depending on hardware. Cloud providers tend to be faster and more accurate but raise privacy and cost concerns.
The audio capture pipeline hooks into system audio devices using Electron’s native capabilities and OS-level APIs. This pipeline feeds raw audio streams concurrently to the selected ASR backends. Transcription results from all backends are aggregated and normalized.
Once transcribed, the text flows into a local-first AI Review Desk, a workspace built into the app for refining and interacting with transcripts. This desk supports:
- Correction modes with diff views to compare raw and edited transcripts
- Multi-threaded chat interfaces that cite transcript references for context
- Structured briefings that summarize content
- Mind maps to visualize relationships between ideas
This set of features helps turn raw, often noisy ASR output into polished, usable text.
UI-wise, DeLive offers a draggable, always-on-top floating caption window. This window can display source text, translations, or both simultaneously, which is handy for live presentations or multilingual environments.
For data persistence, DeLive uses IndexedDB and localStorage, meaning all data is stored locally by default. Users who want cloud backup options can enable syncing with S3 or WebDAV-compatible services.
Beyond the UI, DeLive exposes a local REST API, a real-time WebSocket interface, and notably, an MCP (Model Context Protocol) server. The MCP server lets external AI agents query ongoing transcripts, search transcript history, and trigger correction or annotation actions programmatically. This design transforms the app from a simple transcription tool into an interactive AI skill within larger multi-agent workflows.
What sets DeLive apart technically
DeLive’s biggest technical strength lies in combining multiple ASR backends with a programmable MCP server interface. Most transcription apps focus only on capturing and displaying text. DeLive treats transcription as just one piece of a larger AI workflow puzzle.
The multi-ASR orchestration is non-trivial. Routing system audio to twelve different backends, handling asynchronous responses, and merging results requires careful concurrency control and normalization logic. It balances responsiveness with accuracy by allowing fallback or hybrid transcription strategies.
The AI Review Desk is another highlight. Rather than dumping raw transcripts into a text editor, it integrates features like diff views and multi-threaded conversations tied to transcript segments. This makes the app suitable not just for live captioning but also for post-processing, note-taking, and research.
The floating caption window’s multi-mode display supports use cases ranging from accessibility (closed captions) to live multilingual translation.
Data storage is local-first, which is a deliberate privacy and reliability tradeoff. IndexedDB and localStorage provide offline persistence but have size and performance limitations compared to dedicated databases. The optional S3/WebDAV backup offers a middle ground for users wanting cloud sync without vendor lock-in.
The MCP server implementation is what really positions DeLive as a programmable AI skill. MCP is an emerging protocol designed for multi-agent coordination. By exposing transcripts and correction tools via MCP, DeLive lets other AI agents consume system audio context and orchestrate actions, like automated corrections or context-aware responses, in real time.
This integration is rare in desktop transcription tools. It effectively turns your machine’s audio environment into a live data source for LLM workflows and agentic automation.
Compared to common transcription apps like Otter.ai, Rev, or even local Whisper GUI tools, DeLive’s approach is more extensible and developer-centric. It’s less polished as a consumer product but offers far more programmatic control and integration options.
Quick start
If you want to try DeLive locally, the README provides straightforward commands:
git clone https://github.com/XimilalaXiang/DeLive.git
cd DeLive
npm run install:all
npm run dev
This will clone the repo, install all dependencies, and launch the app in development mode. The app’s UI and features should be accessible once running.
The README also points to a full development guide covering setup, building, and testing if you want to dig deeper or contribute.
Verdict
DeLive is a solid choice for developers and researchers who want a programmable, local-first transcription environment that fits into AI agent workflows. Its multi-ASR backend orchestration and MCP server integration make it more than a simple captioning tool — it’s a platform for building interactive AI skills around system audio.
The tradeoffs are clear: it’s an Electron app, so expect a typical desktop app footprint and some resource usage. Local storage via IndexedDB is convenient but not enterprise-grade. The UI is functional but not flashy.
If your use case demands privacy, offline transcription, and programmability for AI agents, DeLive offers a practical, open-source starting point. It’s not a plug-and-play consumer product but a platform for experimentation and integration.
For anyone working with real-time speech data, agentic AI, or multi-modal workflows, exploring DeLive’s code and architecture is worth your time.
→ GitHub Repo: XimilalaXiang/DeLive ⭐ 174 · TypeScript