Claude Code Routine Connector Keeps Asking for Permission: The Hidden created_via Field

🌏 中文版

TL;DR

Claude Code Routines have an undocumented created_via field that determines connector permission behavior. meta_mcp (created by Claude through MCP tools) prompts for approval on every connector call. http_api (created by the user or API) doesn't. Recreate the routine via the RemoteTrigger API to fix it.

Context

I run over a dozen Claude Code Routines for daily scheduled tasks. One of them — a daily Taiwan stock risk dashboard — needs to call a self-hosted groundlane MCP server to fetch stock data from the FinMind API. This routine was created by Claude during a Claude Code session.

Other routines (like daily-digest-signals) also use the same groundlane connector heavily — over 75 web_fetch calls per run — and never prompt for authorization.

Problem

Every time the stock dashboard routine called groundlane's web_fetch, this prompt appeared:

Allow Claude to use Web fetch (groundlane)?
This connector call requires your approval to proceed.

Scheduled runs that hit this prompt with nobody around to approve would fail with stop_reason=tool_use. Manual runs required clicking "Allow once" each time.

The connector settings page had all groundlane tools set to "Always allow". It made no difference.

Investigation

Ruling out hypothesis 1: first-party vs third-party connector

Initial theory: groundlane is self-hosted (no claude_ai prefix in tool names), so the platform treats it differently from Anthropic-hosted connectors like Exa or Tavily.

Disproved: The daily-digest-signals routine uses the same groundlane connector and its SKILL.md explicitly says "do not switch to Exa/Tavily/others." It never prompts.

Ruling out hypothesis 2: new routine needs first-time authorization

Theory: a newly created routine needs one round of manual approval before auto-approving.

Disproved: The stock routine's first Scheduled run appeared to "succeed," but the run log revealed it never mounted the repo (No sources configured, /home/user was an empty directory). It never reached the groundlane call. The "success" was a false positive.

Finding the real difference

Using the RemoteTrigger API with list and get actions, I pulled the full JSON for both routines and compared field by field:

The groundlane connector configuration was byte-for-byte identical — same connector_uuid, same URL, same empty permitted_tools. The only structural difference was created_via.

Solution

Create a new routine with identical config using the RemoteTrigger API's create action:

# Inside a Claude Code session, use the RemoteTrigger tool
RemoteTrigger({
  action: "create",
  body: {
    name: "Your routine name",
    cron_expression: "30 10 * * 1-5",  // weekdays 6:30 PM GMT+8
    enabled: true,
    job_config: {
      ccr: {
        environment_id: "env_...",
        session_context: {
          model: "claude-sonnet-5",
          sources: [{ git_repository: { url: "https://github.com/..." } }],
          allowed_tools: ["preset:default", "Bash", "Read", ...]
        },
        events: [{ data: { uuid: "...", type: "user",
          message: { content: "Your prompt", role: "user" }
        }}]
      }
    },
    mcp_connections: [
      { connector_uuid: "...", name: "groundlane", url: "https://..." },
      // ... other connectors
    ]
  }
})

The response confirmed created_via: "http_api".

Manually triggering the new routine, groundlane web_fetch returned results immediately with no authorization prompt:

[12:14:29] tool_use mcp__groundlane__web_fetch → FinMind TAIEX
[12:14:31] tool_result: {"ok":true, ...}  ← direct success
[12:14:34] tool_use mcp__groundlane__web_fetch → TAIEX historical data
[12:14:36] tool_result: {"ok":true, ...}  ← success again

Then delete the old routine at https://claude.ai/code/routines.

Root cause

The created_via field records how the routine was created:

• http_api: Created through Claude Code's HTTP API — this includes the web UI, the RemoteTrigger tool, and the /schedule skill. The system treats this as "explicitly authorized by the user," so connector calls proceed without confirmation.
• meta_mcp: Created by Claude through MCP tools during a session. The system treats this as "AI-initiated, user hasn't explicitly approved connector permissions," so every connector call prompts for confirmation.

This field is immutable — set at creation time. Using the update API to change other routine settings does not change created_via. The only fix is to recreate.

Takeaway

Connector permissions in Claude Code Routines don't just depend on the connector's own settings (Always allow / Ask / Deny). They also depend on who created the routine. Letting Claude "create a routine for you" during a session and creating one yourself are two different things in the permission model.

References

• Claude Code Routines documentation
• Model Context Protocol specification
• FinMind Taiwan Stock API (in Chinese)