Most of my working day happens on calls: customer calls through the office PBX, and meetings in Zoom, Teams and Meet. The tools for those two worlds never talk to each other. Softphones just dial. Meeting bots join your calls, upload the audio to someone else's cloud and charge per seat.
So we built AI-Softphone: one desktop app for Windows, macOS and Linux that makes SIP calls, records any meeting on your computer, and turns every conversation into a searchable transcript and a short write-up. It can run 100% locally. It's free and GPL-licensed.
#
What it does
Before and during the call
- A live prompter shows subtitles for both sides, translates the other person line by line, and suggests what to say next based on your instructions (a sales script, an interview guide).
- Four ready assistants: subtitles, translation, objection handling, interview help. All editable.
- Practice mode: replay an old recording and tune your prompter before the real call.
After the call
- Speaker-synced transcripts with click-to-play, full-text search across all conversations, export to text or subtitles.
- Summaries, action items, names mentioned, categories, red flags, and scoring against your own criteria, from 11 editable prompts.
- Ask questions about any past conversation.
Any conversation, one library
- Calls through the built-in SIP softphone (any IP PBX: multiple accounts, transfer, conference).
- Zoom, Teams, Meet or any other app, captured from your computer.
- Imported recordings.
#
Run it fully local
If your conversations can't leave the machine, pair it with local models:
1. Install [Ollama](https://ollama.com) and pull a model, for example`ollama pull llama3.1:8b` .
- In AI-Softphone, open the AI processing settings and pick Whisper or Vosk for speech recognition.
- Point the language model at your local Ollama (or vLLM) endpoint and hit the connection test button.
- Record a meeting or make a call. The transcript and summary are produced on your computer.
Prefer the cloud? Bring your own keys for OpenAI, Claude, Gemini, Mistral, DeepSeek, Groq, Deepgram, AssemblyAI and others. Every call shows what it cost, and a monthly cap s processing before you overspend.
#
For developers and admins #
Stack: C++20, Qt 6 / QML, PJSIP. There's also a headless CLI. #
Integrations:tel: links for click-to-dial from a browser or CRM, a localhost-only REST API (calls, contacts, history, accounts), and webhooks for call start, state change and end. #
No cloud in the middle: data stays in a file you own, passwords go to the OS keychain, and nothing listens on external interfaces. #
Modular: switch off recording, transcription, write-ups, meeting capture, import or integrations independently. With everything off, it's just a softphone.
#
Limitations
- macOS needs Apple silicon.
- No team features yet (central admin, shared storage, reporting). That's the planned Enterprise edition.
#
Try it I'd love feedback, especially from people who run their own PBX or already use a meeting-notes tool: what's missing, what breaks, what would make you switch? A star on GitHub helps a lot too.