plus1

Most AI meeting assistants just record. They transcribe the call and send you a summary after everyone has left. By then, you've probably already done the follow-up work yourself.

We thought about the meetings where you wish you had one more person. A client call where you're talking and also trying to pull up numbers. A team sync where someone says "can someone file that?" and nobody does. plus1 is that extra person you invite along: it sits in the call, listens, and picks up the work so you don't have to context-switch away from the conversation.

I worked on getting plus1 into the meeting and on everything between it hearing something and doing something:

The first version of the agent brain was eager. It would jump in on anything that sounded like a question. In our first test call, it kept interrupting people mid-sentence. We needed to figure out how to get the AI to talk only at the right moments.

So I rewrote the decision layer around a simple question: what would a thoughtful colleague do right now? Most of the time, the answer is nothing. Listen, hold the context, and wait. The confidence gate I built reflects that: every utterance gets classified with a confidence score, and the default is to stay quiet. The agent only speaks when it's clearly been addressed, or when it finishes a task someone asked for.

After several trials, we created these rules:

The "only speak when someone talks to you" rule worked well in a group, but in a call with just one other person, plus1 felt broken. That person is always talking to plus1, but they don't say its name every time. So plus1 checks how many people are in the call, and if it's just one person, it treats everything as directed at it unless it's filler like "one sec." This was a small fix, but it completely changed how the product felt in the most common use case.

plus1 doesn't just answer questions. It does things. The agent brain decides what tool to call, and a single executor runs it. Every tool namespace, whether it's insurance quoting, document generation, browser automation, or email, plugs into the same interface. The brain picks the tool and passes a query; the executor routes it and returns a result the agent can narrate back to the room.

This matters because it means adding a new capability doesn't touch the decision layer at all. You write the tool, register it, and plus1 can use it in its next meeting. It's the same idea behind MCP: tools are separate from the model, and you can swap them in and out without changing how the agent thinks. That's what made it possible for four people to add tools in parallel without stepping on each other.

For the demo, we built tools for insurance quoting (with a deterministic pricing engine, so every number is auditable and explainable), VIN lookup through a real NHTSA API, PDF generation, email, underwriting, and browser automation. But the point was never the insurance. It was that plus1 could pick up any job you give it, as long as someone writes the tool.

The dashboard shows a live transcript, how plus1 sorted each line, the tools it called, and cards for anything it made. This was one of the first things I built, before most of the tools. If you can't see why an AI did something, it's hard to trust it, and trust matters a lot more when the AI is talking out loud in a room with your client.

Behind it, Playwright runs a Chrome window that joins the Meet. A HeyGen avatar and an ElevenLabs voice act as its camera and mic, OpenAI Realtime handles the speech-to-speech path, Gemini handles the decision classification, and Browserbase runs a cloud browser it can share with the call.