Real-time transcription
Low-latency streaming speech-to-text from Deepgram's nova-3 model, for both your mic and system audio.
Open source · Real-time meeting copilot
Live meeting answers, grounded in your code and context.
Echolect is an open-source desktop copilot that listens, transcribes, and answers while the call is still happening — using your live transcript, project memory, what's on your screen, and your source code.
Transcript
DevonQuick standup — where are we on the provider refactor?
YouLive and batch both go through one session factory now.
DevonDoes research still get read access to the repo?
Answer
Yes — research runs in its own session with the same meeting directory and read-only repo scope, then buffers its findings into your next answer so the live session stays fast.
Screenshot #1
analyzed
Research
Deepgram nova-3 streams diarization inline — multi-speaker labels need no extra setup.
Ask the assistant…
What it does
A complete live-meeting loop — not just a recorder.
Everything you need to follow a conversation and act inside it, grounded in the context that actually matters.
Speaker naming
Speakers are separated and named automatically from the conversation as it unfolds. Rename anyone — your edits stick across meetings.
On-demand assist
An always-on-top overlay with Answer, Suggest, Ask back, and Explain intents — fired by click or hotkey, or just type a question.
Grounded in context
Every meeting is pre-loaded with your personal and project context, prior summaries, and — for code work — read-only access to the repo.
Screenshot analysis
Grab the screen with one key; Echolect reads it — slides, errors, diagrams, code — and folds that understanding into your next answer.
Live running summary
A summary panel that keeps refreshing as the meeting progresses, so you always have the gist at a glance.
Web research sidecar
Ask a question and a separate session searches the web and reads project files, feeding findings back to the live assistant.
Claude or Codex
Runs on the Claude CLI or the OpenAI Codex CLI — whichever you have installed. Switch providers and models in Settings.
Memory that compounds
When a meeting ends, its summary is written and rolled into the project's growing context — so the next meeting starts sharper.
Why it's different
Inspired by Parakeet AI — built to go further.
Live meeting copilots like Parakeet AI showed how useful real-time, in-call assistance can be. Echolect takes the idea further: it's open source, runs on your own machine, and grounds every answer in your local knowledge base — source code, project context, and past meetings — not just the call transcript.
Typical bots
- Join the call as another participant.
- Transcribe and summarize after the fact.
- Answer only from the call transcript and generic model knowledge.
- Treat your repository and local files as out of band.
- Keep your memory locked inside their hosted app.
Echolect
- Runs as a desktop overlay and answers while the conversation is happening.
- Grounded in your project context, prior meetings, screen, and source code.
- Reads your repo through controlled, read-only tools when source is the truth.
- Stores everything as readable local files that you own.
- Each summary feeds the next meeting — memory that compounds, locally.
How it works
A fast live path, with side paths that never block it.
01
Listen
Mic and system audio stream to Deepgram, get speaker-labeled and de-duplicated, and land in a durable transcript.
02
Ask
An intent or a typed question sends only the new transcript delta into a warm session that already holds your context and code scope.
03
Add context
Screenshots, web research, and the scribe's names and summary buffer in and fold into your next answer — never blocking the live thread.
04
Finish
Speakers are reconciled, a summary is written, and it rolls into the project's context — ready for next time.
Want the full picture? See the architecture →
Download Echolect
Free and open source. Grab the latest build for your OS, then add your keys in Settings on first launch.
Linux Available
.AppImage (portable, runs anywhere) or .deb for Debian / Ubuntu.
Windows Available
NSIS installer (.exe) — per-user, no admin prompt, with a desktop shortcut.
You'll also need:
A Deepgram API key — for transcription
The Claude or OpenAI Codex CLI on your PATH — for the AI
Roadmap
What's coming in v2.
Voice control
A wake phrase plus spoken commands — "answer that", "summarize", "screenshot" — so it works hands-free while you share your screen.
Multilingual support
Transcription, speaker diarization, and assistance in languages beyond English, selectable per meeting.
Runtime prompt sets
Save named bundles of the assistant's prompts — Technical interview, Client call, Standup — and switch between them mid-meeting.
Hidden overlay on Windows
Keep the overlay visible to you but invisible to screen-sharing and recording, so it stays private during calls.
OpenRouter key support
More model access without forcing everyone into the same local CLI setup.
Shaped by the community
It's open source — issues, ideas, and pull requests on GitHub steer what ships next.