For a browser-based transcription server, start with Scriberr or Whishper. For live meeting notes on your own computer, compare Anarlog and Meetily. Vibe fits desktop transcription and local AI analysis; WhisperX is a component for a custom pipeline.
“Self-hosted AI notetaker” can mean three different things: an app that runs models on your laptop, a web service on your server, or a transcription engine you integrate yourself. Choose the deployment first. Installing a local desktop app does not automatically give a team a shared server with accounts and access controls.
We build Anarlog. This comparison uses current official repositories and documentation; it does not report a hardware or accuracy benchmark.
Self-hosted AI notetaker comparison
| Tool | Deployment | Local transcription | Summaries | License and cost |
|---|---|---|---|---|
| Anarlog | Desktop on macOS, Windows, Linux; enterprise deployment evaluated separately | Built-in speech paths on supported Apple Silicon; compatible local server on other platforms | Local model or selected cloud provider | MIT community app; free local/BYOK, Pro $14/month |
| Meetily Community | Desktop on Windows/macOS; Linux source build | Whisper or Parakeet | Ollama or BYOK | MIT; free Community, commercial Pro separate |
| Scriberr | Browser UI on a self-hosted server; Docker or Homebrew | Whisper, Parakeet, Canary; diarization | Ollama or OpenAI-compatible provider | MIT; free software |
| Vibe | Desktop on macOS, Windows, Linux; HTTP API | Whisper, Nemotron, Parakeet | Ollama or Claude API | MIT; free software |
| Whishper | Docker web service | FasterWhisper | No meeting-summary workflow advertised | AGPL-3.0; free software |
| WhisperX | Python library / CLI | Whisper-based transcription and alignment | Add your own | BSD-2-Clause; free software |
Free software still needs hardware, storage, maintenance, and downloaded models. Optional cloud AI can add usage charges. Check the license of the application and the models you deploy.
Local desktop apps for meeting notes
1. Anarlog: notes, transcript, and summary in one workspace
Anarlog captures microphone and system audio without a meeting bot. You can write during the call, then combine your notes and transcript in a templated summary. Canonical meeting data lives in local SQLite, with recordings and attachments stored separately and portable exports when needed.
Setup: choose the installer for your platform, then configure transcription and Intelligence separately. The models and providers guide documents each path. Built-in Parakeet Streaming/Batch and whisper.cpp .bin transcription require supported Apple Silicon. Apple Speech also requires compatible macOS 26 system assets. Windows, Linux, and Intel Macs can use a Deepgram-compatible local server or a configured hosted provider.
For local summaries, prepare an Ollama or LM Studio model and select it under Intelligence. A local speech model with a cloud summary model is a mixed data path. The offline Mac setup walks through testing both stages.
Limits and cost: the community desktop app is MIT-licensed and free for local/BYOK use; a paid API provider can charge separately. Pro is $14/month or $140/year and adds hosted AI, encrypted CloudSync, sharing, and calendar-connected services. Sharing has a distinct server-readable data path, described on the security page.
The MIT desktop app and a customer-hosted enterprise stack are different distributions. The enterprise page labels customer-hosted capture and data-plane work as early-partner evaluations; that is a scoped commercial deployment discussion, not a free Docker install for this desktop app.
2. Meetily: local meeting capture and summaries
Meetily's repository screenshot shows the transcript and generated meeting record. Source: Meetily.
Meetily Community captures microphone and system audio, transcribes with local Whisper or Parakeet, and generates summaries with Ollama or a selected API provider. It also documents audio import. It is a desktop application, not a shared browser workspace.
Setup: use Windows or macOS releases; Linux requires a source build. Acceleration depends on the package: macOS uses Metal/CoreML, packaged Windows Whisper uses Vulkan, and CUDA requires an appropriately configured source build. Do not assume an NVIDIA card makes the standard installer select CUDA automatically.
Edition boundary: Community is free under MIT. The current Pro comparison lists speaker identification, custom templates, advanced exports, and assistant/script integrations in commercially licensed Pro. It lists $10/user/month billed annually, alongside a $25 regular monthly price. Confirm the edition and deployment offer before budgeting for a team.
Choose Community for a local capture-to-summary workflow. A cloud summary provider receives transcript text; choose local Ollama when that text must stay on your own machine.
3. Vibe: desktop transcription with optional local analysis
Vibe's desktop interface. Source: Vibe.
Vibe's current repository documents file and batch transcription, microphone/system-audio capture, speaker diarization, an HTTP API, and summaries through Ollama or the Claude API. It supports Whisper, Nemotron, and Parakeet models. Describing it as transcription-only misses its current capabilities.
Setup: install the macOS, Windows, or Linux desktop build and download a supported model. Exports include TXT, SRT, VTT, JSON, PDF, and DOCX. Use Ollama for local analysis; the Claude route sends text to a cloud service.
Deployment fit: this is an MIT-licensed desktop app with API access. An HTTP endpoint is useful for automation, but it does not by itself establish a supported multi-user team service. If you need a shared browser workflow, compare the server options below.
Browser-based services you can deploy
4. Scriberr: recordings, transcripts, and AI in a web UI
Scriberr's recording library. Source: Scriberr.
Scriberr is an MIT-licensed server application with a browser UI. It documents uploads, browser recording, annotations, diarization, APIs, and Ollama or OpenAI-compatible summaries. Transcription options include Whisper and NVIDIA Parakeet/Canary.
Setup: follow the repository's CPU or NVIDIA Docker configuration, or its macOS/Linux Homebrew installation. Persist both application data and model environments; the README distinguishes these volumes. Choose the GPU image for your hardware.
Limits: this is a service you operate. Use the documented HTTPS/reverse-proxy configuration and evaluate access separation before making it a shared team workspace. Browser recording is not automatic system-audio capture from every conferencing app.
The maintainer now says development is resuming, superseding the old pause announcement. Check releases and issues for your version.
5. Whishper: transcription, translation, and subtitle editing
Whishper's upload dialog. Source: Whishper Docs.
Whishper is a Docker-based transcription and subtitling service under AGPL-3.0. Its documented stack includes FasterWhisper, a web UI/backend, MongoDB, optional LibreTranslate, and Nginx. It supports file/URL input, CPU or NVIDIA processing, and TXT, JSON, VTT, and SRT exports.
Its subtitle editor lets you change text and segments in the browser. This is more than a raw transcript downloader, but the project does not advertise a meeting-summary workspace.
Limits: user authentication and browser audio recording remain roadmap items in the main README. A reachable URL is therefore not enough to call the default deployment a private team service. Plan access control before sharing it.
The maintainer states that the main branch will receive no further releases while a v4 rewrite is underway. Evaluate the current branch's maintenance status and license before adopting it for a long-lived service.
A transcription engine for a custom stack
6. WhisperX: alignment and speaker labels
WhisperX is a CLI and Python library. Source: WhisperX on GitHub.
WhisperX combines Whisper-based transcription, word alignment, and optional pyannote speaker diarization. The current project license is BSD-2-Clause.
Setup: install with pip install whisperx or uvx whisperx and follow the repository's dependency instructions. CUDA is an acceleration option; the README explicitly documents CPU use with --device cpu --compute_type int8. Model size, memory, and recording length affect performance, so test your own workload.
Diarization requires a Hugging Face token and acceptance of the relevant model terms. Alignment coverage depends on language-specific models. Overlapping speech and speaker assignment still need review.
Deployment fit: WhisperX supplies the processing pipeline. You add capture, storage, access control, a user interface, and summarization if you need a complete meeting service. It is not the only project here with speaker labels: Scriberr and Vibe also document diarization.
Which deployment should you choose?
- Live notes on one computer: Anarlog combines editable notes with capture and summaries; Meetily Community offers local capture and summaries.
- Desktop files and local AI analysis: compare Vibe's batch, export, and Ollama workflows.
- An upload/recording server with AI: evaluate Scriberr and its current deployment instructions.
- A subtitle-focused web service: evaluate Whishper, including its access-control and maintenance limits.
- A custom processing backend: use WhisperX when you want to integrate alignment and diarization yourself.
- Company-managed infrastructure and policy: distinguish community tools from commercial enterprise deployments. Anarlog's enterprise evaluation is a separate path.
Before choosing, process the same representative recording with your intended setup. Check transcription, speaker attribution, summary decisions, export, and whether either AI stage calls a cloud provider. Prepare models before testing offline. For a server, also test user access and recovery from a backup; one working upload does not establish those behaviors.
For broader file and meeting choices, use the transcription guide hub or open-source meeting transcription comparison. To try Anarlog's desktop workflow, choose the matching installer.