Weave In: Bringing Everyone’s Ideas into the Conversation

Weave In is an agent-native meeting app built around collaboration in group meetings. We focus on the problems that groupthink can create: when someone proposes a plan, others may agree before discussing it fully, leaving different perspectives and concerns unspoken. Meetings also present another difficulty. In a lively discussion, someone may have an idea but struggle to find an opening. By the time they can speak, they may have forgotten the idea, or the conversation may have moved on.

Weave In supports critical thinking in collaboration: maintaining independent judgment, asking questions, and examining a proposal’s assumptions and evidence together. It gives valuable ideas a place to be captured and clarified before they enter the shared discussion. That is why Weave In takes the form of a meeting app, offering support while the conversation is happening and ideas can still shape a decision.

The full experience follows an agent-native design. Instead of starting with a chatbot inside a meeting app, we reverse the starting point: the meeting app runs in a browser that an agent can operate. Users can bring their familiar agent into a meeting, let it understand the discussion and carry out tasks, and bring its work back to the shared space. Weave In also works in ordinary browsers. In browsers that support WebMCP, such as the browser built into Codex, its tools connect a personal agent to the meeting.

Three kinds of agents serve different needs in this design.

The first is Omni, the team agent. It follows the team’s shared goals and helps draw attention to unresolved questions and departures from the agenda. For example, if the group is about to approve a proposal without addressing an earlier concern, Omni can bring that question back into the discussion. If the conversation drifts from its original goal, it can prompt participants to decide whether to return to the topic or adjust the agenda. Omni quietly prepares a question, then waits for a participant to approve it with a button or, when captions are enabled, say “Omni, go ahead.” It then reads the question aloud during a gap in the discussion. Omni helps the team stay attentive to the issues; participants make the decisions.

The second is Muse, a personal agent for live conversation. Powered by GPT-Live-1, it lets users have a private voice side chat during a meeting. When an idea is still uncertain, someone’s explanation raises a question, or a formula or chart on a shared screen is hard to follow, users can discuss it with Muse first. They can build their understanding, organize their thoughts, and decide how to bring them back to the meeting. Entering a private Live conversation pauses the meeting microphone and public captions, restoring their previous state when it ends. Users can also type or dictate an editable draft. Private discussions stay in the personal space. Users can choose a completed Muse reply and have just that reply read aloud to everyone.

The third is the personal agent a user already works with, such as Codex. It handles asynchronous tasks that can run in the background, including looking up information, comparing approaches, checking code, and preparing diagrams. These tasks can progress alongside the meeting, giving participants further analysis and verification while the discussion continues.

To connect Codex’s work to the meeting, we provide meeting state and actions through WebMCP. Within its authorized scope, Codex can read participants’ live transcripts, public chat, screen shares, and whiteboard content. Following the user’s instructions, it can publish completed work to public chat or place it on the shared whiteboard. This gives the agent the context of the ongoing discussion and lets everyone build on its results.

Ideas in a meeting also need a place where people and agents can examine, discuss, and revise them together. The shared whiteboard, built on Excalidraw, keeps workflows, architectures, and relationships on one canvas. Participants can point out problems, add details, or draw their own ideas. We integrate the whiteboard into a meeting app environment we control so that agents can use WebMCP’s structured tools to read nodes, text, and connections as structured context, then add, modify, or connect elements. This gives agents a clear interface for understanding the current state, making precise changes to shared work, and continuing from a person’s latest edits.

To make diagram creation easier for agents, Codex can describe a workflow or architecture in Mermaid. Weave In converts Mermaid flowchart and graph syntax into editable whiteboard nodes, text, and connections. Agents create diagrams in a familiar text format, while people can move nodes, edit labels, and adjust arrows directly on the canvas. Agents can still read and edit the resulting structure, so both can keep working on the same artifact. Muse can also operate the whiteboard when the user directly asks it to. To publish text to Room, the user first enables publishing permission, then asks Muse to share the specified content.

For example, a team is discussing a new feature. One member has concerns about the implementation but cannot find a moment to speak. They first clarify the issue with Muse, then ask Codex to inspect the relevant code. Once Codex finishes, it follows their instructions to share its analysis and a flowchart with the meeting. Everyone can then adjust the approach directly on the whiteboard. If the discussion drifts from its original goal, Omni prompts the team to revisit the agenda after receiving a participant’s approval. An idea that might otherwise have gone unspoken gets a chance to become part of the shared discussion.

The implementation brings together group meetings, live transcripts, and a shared whiteboard with Muse’s live conversations, WebMCP tools for reading state and taking actions, and workflows for team reminders and asynchronous sharing. Codex is also a development tool, helping us implement features, inspect tool interfaces, and verify the complete experience from reading the meeting context to delivering editable work.

Weave In connects live conversation, private thinking, and asynchronous work so that everyone can contribute to the shared discussion in a way that works for them.