Skip to content

[Coding Plan] MCP tools (web_search_prime / web_reader) hit 429 rate limit aggressively — thresholds undocumented, unrelated to quota #151

Description

@Enominera

Summary

GLM Coding Plan 的官方 MCP 工具(web_search_prime / web_reader)触发 429 限速的阈值极低且数值未公开,且与套餐额度无关,付费用户难以将其纳入日常工作流。

Environment

  • GLM Coding Plan(付费订阅)
  • MCP endpoints: https://open.bigmodel.cn/api/mcp/web_search_prime/mcp, https://open.bigmodel.cn/api/mcp/web_reader/mcp
  • Client: Claude Code (user-level remote HTTP MCP)

Observations (measured, 2026-09-12, single client, single account)

  1. 一天内约 8 次 MCP 调用,其中约一半返回:
    {"error":{"code":"1302","message":"您的账户已达到速率限制,请您控制请求频率"}}
  2. 429 后 20+ 分钟不恢复(限速窗口看起来是小时级,而非分钟级 QPM)。
  3. 与套餐额度无关的实证:控制台显示 5 小时额度仍剩 58% 时持续 429;同时段模型 API 调用完全正常。即 MCP 限速是独立于额度池的闸门。
  4. web_search_primeweb_reader 疑似共享同一限速池(交叉触发)。
  5. 已排除多客户端并发:本机仅一个客户端在调用。

Requests

  1. 公开 MCP 限速数值(QPM / 窗口长度)——docs 的 coding-plan/usage-notesfaq 均未披露,用户无从规避。
  2. 放宽阈值或按套餐分级:付费 Coding Plan 用户将 MCP 作为 agent 检索工作流的一环,当前阈值下基本不可用(精搜节律一天几次都会撞)。
  3. 429 报文语义区分:限速(速率)与额度耗尽(quota)是两回事,建议报文/文档明确区分,避免误导排查方向。

Filed by an AI agent (Claude Code via cc-connect) on behalf of the account owner, with their explicit request. Measurements above are from the owner's paid account.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions