Condux

让AI代理能够从生产环境中实际出错的环节进行调试。连接到Condux项目后,它可以使用仪表板自带的筛选器搜索项目问题,打开带有最新堆栈跟踪的问题,并翻阅过去的发生记录。令牌默认限定于单个项目,且只读。

托管 MCP 服务器

npx add-mcp 'https://app.condux.ai/api/mcp'

可安装到 Claude Code、Codex、Cursor 等客户端

文档

Give your AI agent your production errors

Condux exposes a Model Context Protocol endpoint, so Claude, Cursor or any MCP client can list issues, read a de-minified stack trace and page through real occurrences. A token is read-only unless you grant it triage, it burns no fix allowance either way, and it resolves to exactly one project.

Get started free

Debugging with an agent usually means pasting a stack trace into a chat window. An MCP endpoint removes that step: the agent queries the errors themselves, in your editor, against what actually broke in production. The interesting question is what you had to hand over to get that, and the answer here is one project, and only the access you chose when you minted the token.

Connect a client

Mint a token in the project's MCP tab, then paste the config your client generated it for.

{
  "mcpServers": {
    "condux": {
      "url": "https://app.condux.ai/api/mcp",
      "headers": { "Authorization": "Bearer condux_mcp_..." }
    }
  }
}

What the agent can do

  • list_issues Search and filter issues with the same query grammar as the dashboard, by status, level or assignee.
  • get_issue Read one issue with its latest event, including stack frames resolved through your uploaded source maps.
  • list_issue_events Page through real occurrences of an issue, so the agent sees the pattern rather than one sample.

What it deliberately cannot do

The scope is the feature. An agent credential is only as safe as the smallest thing it can reach.

  • Writes only when you grant them A token is read-only unless you mint it with triage, which adds two tools: set an issue's status and add a note. No tool changes a project or a setting, or requests a fix, and the capability is fixed when the token is minted.
  • One project per token A token resolves to a single project. It is not an account credential and it cannot reach a second project.
  • No allowance to burn Reading errors through MCP costs no Conductor fix runs, so an exploring agent cannot spend your month.
  • Already scrubbed Events are scrubbed of PII at ingest, so the agent sees exactly what the dashboard shows and nothing more.

Frequently asked questions

Can the agent change anything in Condux?

Only if you grant it. A read token has three read tools and nothing else. A triage token adds two: resolve, ignore or reopen an issue, and add a note, which is attributed to the token. No tool changes a project or a setting, or requests a fix.

Does using MCP consume my AI fix allowance?

No. The allowance covers Conductor runs, which open draft pull requests. Reading issues over MCP is free and unmetered.

Which clients work?

Any MCP client that can talk to a remote HTTP server and send a custom header, which includes Claude and Cursor. Condux speaks JSON-RPC 2.0 over Streamable HTTP.

Does it work when I self-host?

Yes. The endpoint is part of the control plane, so a self-hosted Condux serves it from your own deployment with no extra service to run.