HaiAI123

Curated Global AI Tools Directory

Industry update
Flat

Local Agent Control Protocol for Coordinating Multiple Web AI Models

11/15/2026 — 11/15, 12:16·1 sources (all)·1 reports (all)

Key takeaways

  • The author reviews the control protocol behind a local daemon that coordinates multiple web-based AI models. The design is shaped by MV3 ext…

Background & analysis

On 2026-11-15 at 12:02, a design retrospective published on Juejin reviewed the control-plane protocol for a local agent that directs multiple web-based AIs, focusing on the design trade-offs behind a multi-model group chat setup. In an earlier post, the author covered the data plane: how to type a question into a webpage input box and how to tell that a response is finished in a world where the web AIs expose no finish_reason, by pulling five web AIs into a single group chat for multi-Agent orchestration without any API key. This article turns to the other half of the problem, the control plane, which decides who is allowed to make whom work and how failures can be recovered from on their own. The author notes that when the caller is a local agent such as Codex or Claude Code rather than a human, the control plane is what determines whether the whole approach is usable.

The protocol spans four hops, two more layers than the simpler idea of letting the extension open an HTTP port directly, and each extra layer was forced by hard constraints rather than design taste. A service worker in an MV3 extension cannot listen on a port. It can only act as a client that opens fetch or WebSocket connections, never as a server that accepts them, so an additional local daemon running on 127.0.0.1 is required and the CLI connects through it instead of reaching the extension directly. Requests are paired with their results by request id, timeouts are treated as local failures, and error codes carry a hint so that the agent can recover on its own. The port is not hardcoded but exposed through the agentControlPort setting, and the migration logic keeps LEGACY_DEFAULT_PORT: if an older store at version < 7 holds the old default port 19826, it is switched to 19305 on read, so that upgraded users avoid the hardest-to-debug state of a daemon that is up while the extension cannot connect. The author also flags port migration and the need to allow WebSocket in the extension page CSP as two pitfalls hit in practice. Only this one report is available so far, and there is no further update on how the protocol has evolved since.

AI-generated from 1 reports · updated 2 days ago

updream
24-hour heatHeat index 25 · peak 49 · 23h ago
24 hours agonow

Reports on this story headlines open the original

Nov 15
  1. The author walks through the control plane of a multi-model chat setup: why it needs a local daemon, and how two MV3 constraints — an extension cannot listen on a port, and its service worker is killed when idle — shaped the protocol. Requests are matched by id, timeouts fail in place, and error codes carry hints so an agent can recover on its own. Port-migration handling and the extension page's CSP allowance for WebSocket are flagged as the two easy-to-miss pitfalls.

    掘金AI score 54

Other stories people are talking about

How is heat calculated?About the method

Heat counts how many independent sources covered a story in the last 48 hours: one source counts once no matter how many posts it published, decaying with a 24-hour half-life. What ranks first is what many people are talking about.

This page aggregates public feeds. Headlines and summaries are machine-organized and remain the property of the original authors; verify important facts at the source. If you believe a headline or summary infringes your rights, email the contact address in the footer with the page URL and basis of your claim, and we will remove or replace it after review.

Surge
Discussion rising fast
New
First report within 6 hours
Rising
Still gathering discussion

Back to AI industry updates →