The Round Table

Context

Comparing model responses usually means moving long answers between separate windows. That becomes awkward when an answer contains tables, source material, or several rounds of reasoning. I wanted one room where both models could see the same conversation and respond to each other without repeated copying and pasting.

Approach

The tool keeps a shared transcript for a human moderator and two configurable model participants. Moderated mode leaves turn-taking to the user. Guided mode runs a selected protocol such as refinement, point-counterpoint, or red-team review. Autonomous mode permits a fixed number of exchanges and always exposes a stop control.

What happened

The current version supports useful three-way conversations while keeping the user in control of direction and token use. Discussions, model choices, and optional remembered credentials stay in that browser rather than a site database.

Working with AI

Codex helped turn the initial conversation concept into explicit interaction modes, provider adapters, interruption controls, and storage boundaries. The important product decisions came from using the tool: long responses needed continuation support, model selection needed to come from the provider, and autonomous exchanges needed an obvious emergency stop.

Key details

• Moderated, Guided, and bounded Autonomous modes share one transcript.
• OpenAI and Anthropic model lists are loaded from the user’s own provider credentials and filtered to conversation-capable models.
• Discussions, source material, model defaults, and optional remembered credentials are stored only in that browser.
• Every running exchange can be stopped, and discussions can be permanently deleted from the device.

Tools and method

The Round Table is a native Next.js route. Browser code stores discussions and preferences locally, while same-origin route handlers relay individual requests to OpenAI or Anthropic and stream normalized response events back to the room. API keys are never written to a server-side database.

Lessons learned

The useful distinction is not simply asking two models the same question. A shared transcript lets one model test, refine, or disagree with the other, and structured protocols make that exchange easier to direct without manually repeating context.

Limitations

Browser-local discussions and remembered keys do not sync between devices. Clearing site data removes them. Provider limits, model availability, and API costs still apply, and autonomous mode should remain bounded rather than treated as an unattended research system.

Next step

Use the native site version in ordinary discussions, then decide whether an explicit export and import format would be useful enough to add without introducing accounts or cloud storage.