The problem
The meeting helper runs on my screen during client calls for Salesforce work at a consultancy, and it is live.
How it works
In a call, the answer you need is in the org, in an email or in a file, and there is no time to look.
The helper works in levels. The quick level answers in about two seconds: what was just asked, a line I can say, the key facts, the sources and any commitments. The grounded level sends four read-only agents to the org, my local files, email and the web, and merges their findings into one card with citations. A screen level captures the shared screen every two seconds when it changes; if a Salesforce record or an error shows up, the org agent checks it. A general level, about ninety seconds, says what the meeting is about and what is still missing.
- A line to say in about two seconds
- Read-only agents add a cited card
- Never types, clicks, speaks or sends
Outcome
Nothing in it writes: no keyboard, mouse, sound, email or chat. Commitments are tracked as who, what and when.
Stack
- claude-code
- ai-agents
- speech-to-text
- read-only-agents
Code
github.com/HenryLop/live-call-coach
Transcript
During client calls, my screen shows cards as people talk. The quick level answers in about two seconds: what was asked, a line I can say, and the key facts. A deeper level sends four read-only agents to the org, my files, email and the web, and merges a card with citations. When a Salesforce record or error appears on the shared screen, the org agent checks it. It never types, clicks, speaks or sends anything, and it tracks who committed to what, by when.
How this was made
Voiceover is text-to-speech (Microsoft Edge neural voice). Background art is AI-generated (Gemini). Titles and captions are set in code, not generated. The build itself was made with Claude Code.