Corey James- 13meeting actions with exact schemas, each naming the acting employee
- 30 minTeams media soak from Azure to Linux
- 0receive stalls through the soak and a forced mid-call reconnect
Answering calls like a colleague
Helix exists because I wanted one AI colleague that shows up where people already work, under an identity the organisation governs, and behaves in a call and a meeting the way a colleague does. People call the digital employee in Teams and it answers. Call audio crosses from a .NET 8 service on the Microsoft real-time media SDK to a Rust voice engine on my GPUs. When someone talks over the agent, it stops.
A Teams call is one more implementation of the engine's sealed audio interface, so the same conversation engine runs on a desk microphone and in Teams. See A self-hosted colleague in a Teams call.
Acting in meetings, and refusing to fake it
In meetings, the agent acts through an MCP server whose actions each name the acting employee. The second-generation set has exact schemas: chat as the employee, a room-state view of who is present and whose hand is up, calling on raised hands, presenting, and sharing a screen. Each action is exercised in a real meeting, and those scripts leave the meeting and clean up even when an assertion trips.
Where Teams has no supported path, the action returns a named refusal. Raising its own hand is one; see A meeting attendee that will not fake it.
Reporting a hand as raised from local state would be a lie nobody in the meeting could see through.
Moving presentation to the meeting stage
PowerPoint Live exposes no control or event hooks. I gave it up and paid for a rewrite: presentation moved to an installed Teams meeting-stage app built on TeamsJS with an authenticated relay, and the agent drives the deck on the stage itself.
No control or event hooks an agent could use.
The agent drives the deck on the stage, through an authenticated relay.
Sharing a screen people can read
Screen share comes from a headless Linux desktop with a browser and a terminal, captured with FFmpeg, framed as H.264 and published through the media service. Video crosses Linux, Windows and Teams under one typed frame contract, and terminal text appears at a speed people can read.
Delegating work from a Teams tab
Delegated work starts from a Teams personal tab in React with Fluent UI, signed in through Teams single sign-on with MSAL and PKCE, and ownership of each request comes from that sign-in. The identity behind all of it is in Solving the digital-employee identity.
I can make an AI colleague behave properly inside Microsoft Teams, answering, yielding the floor, presenting and sharing, with every action attributed to one governed identity. Teams and its SDKs are Microsoft's; the media service, meeting tools, stage app and relay are mine.