Fix Scheduled Routines MCP 'Tool Call Requires Approval' Error on Claude.ai with Custom and First-Party Connectors (May 2026 Regression)

If routines continue to fail after the platform fix, the issue may be stale permission state from the regression window. Remove and re-add the MCP connector to the routine to force a fresh permission grant with proper bind-time approval. This resolves lingering state from the period when permitted_tools was not being honored as a standing grant.

1. In claude.ai/code, navigate to the affected routine's settings
2. Remove the MCP connector from the routine and save
3. Re-add the same MCP connector to the routine
4. Confirm the 'no approval needed' banner appears for the connector's permitted tools
5. Manually trigger the routine via 'Run now' to verify the fix

# If using the API to manage routines, GET then PATCH to force re-grant:

# First, GET the current configuration:

curl -s -H "Authorization: Bearer $CLAUDE_API_KEY" "https://api.claude.ai/v1/code/triggers/<routine_id>" 2>/dev/null | python3 -c "import json,sys; d=json.load(sys.stdin); print(json.dumps(d.get('mcp_connections',[]), indent=2))"

# Then PATCH with modified mcp_connections to force re-grant:

# curl -s -X PATCH -H "Authorization: Bearer $CLAUDE_API_KEY" -H "Content-Type: application/json" \

#   -d '{"mcp_connections":[...]}' "https://api.claude.ai/v1/code/triggers/<routine_id>"

{
  "mcp_connections": [{
    "connector_uuid": "<uuid>",
    "name": "Slack",
    "permitted_tools": ["slack-post-message", "slack-list-channels"],
    "url": "https://mcp.claude.ai/connectors/slack"
  }]
}

While the routine surface was broken, the same MCP connectors worked perfectly from local Claude Code CLI sessions against identical OAuth grants. For production-critical MCP automations, maintain a local CLI-based cron fallback that can be activated when platform-side routine issues are detected. This pattern also provides resilience against future platform regressions.

1. Set up a local cron job or launchd/systemd timer that invokes Claude Code CLI with the same MCP connector and task prompt
2. Use the --print flag for non-interactive execution: claude --print 'Call mcp__connector__tool with parameters X and reply ok'
3. Monitor both the platform routine and the local cron via a health check that verifies output landed in the destination connector
4. When platform routine fails (output missing), the local cron continues to provide coverage

# Example: local CLI cron invocation with MCP connector (test interactively first):

claude --print 'Using the Slack MCP connector, post "Daily health check passed" to #alerts channel' 2>&1

# Check that local CLI has the same MCP connector configured:

claude mcp list 2>/dev/null || python3 -c "import json; d=json.load(open('$HOME/.claude/mcp.json')); print(list(d.get('mcpServers',{}).keys()))"

# Verify the specific tool is callable (replace placeholders):

claude --print 'Call mcp__<connector>__<tool> with {{"param": "value"}} and reply with the result' 2>&1 | head -20

{
  "mcpServers": {
    "claude_ai_slack": {
      "type": "remote",
      "url": "https://mcp.claude.ai/connectors/slack"
    }
  }
}

An AI agent or developer encountering 'MCP tool call requires approval' needs to determine whether this is the routine-surface regression (platform bug, server-side fix) or a different MCP permission issue. This diagnostic solution provides the checks to isolate the root cause before applying workarounds. The decision tree: (a) If error occurs ONLY in scheduled routines AND MCP works from CLI → this is the routine-surface bug; (b) If error occurs in BOTH routines and interactive CLI → different issue (likely OAuth/permission config).

1. Check the execution surface: is this error occurring in a scheduled routine (claude.ai/code → Routines) or during interactive use (CLI, chat, Desktop)?
2. If in a routine: check last_fired_at — did the routine fire but produce no output? This confirms the bug pattern (routine executes, MCP calls silently fail)
3. If interactive: this is NOT the routine-surface bug — the error here indicates a genuine permission/approval configuration issue that requires user interaction
4. Cross-check: run the same MCP connector from local CLI. If CLI works but routine doesn't, this confirms the routine-surface regression

# AI AGENT DECISION TREE — run these checks in order to identify the root cause:

# Step A: Determine execution context

# If you can run this echo command, you are in an interactive context (CLI/chat), not a routine:

echo 'INTERACTIVE_CONTEXT' 2>/dev/null && echo '→ This is interactive. If MCP works here but not in routines, proceed to Step B.'

# Step B: Verify MCP works from CLI

claude mcp list 2>/dev/null || python3 -c "import json; d=json.load(open('$HOME/.claude/mcp.json')); print(list(d.get('mcpServers',{}).keys()))"

# Step C: Test a tool call from CLI

claude --print 'Call mcp__<connector>__<tool> with minimal params and reply with the result' 2>&1 | grep -v 'Streamable HTTP error' | head -5

# Step D: If Steps B-C succeed but routine fails → confirmed routine-surface regression

Anthropic deployed a server-side fix within ~10 hours of acknowledging the issue on May 21, 2026 (first wave ~15:14 UTC, full resolution ~19:57 UTC). No client-side update is required — the fix is applied at the claude.ai platform level. Routines should resume working automatically. Staff member omid-ant confirmed resolution. If issues persist after the fix window, restart Claude Desktop and re-test.

1. Confirm your routines are now executing: check claude.ai/code → Routines → last_fired_at shows recent timestamps post-2026-05-21T20:00Z
2. Verify tool execution: check the destination connector (e.g., Slack channel, Google Calendar) for expected output from the most recent routine run
3. If routines still fail post-fix: try removing and re-adding the MCP connector to the routine to force a fresh permission grant (see Solution 4)
4. If the issue persists: file a new bug report referencing #61015 and the specific connector/error observed

# Diagnostic: check if the routine ran recently and what connectors are attached

# This helps an AI agent distinguish this bug from other MCP failures

curl -s -H "Authorization: Bearer $CLAUDE_API_KEY" "https://api.claude.ai/v1/code/triggers" 2>/dev/null | python3 -c "
import json, sys
triggers = json.load(sys.stdin).get('data', [])
for t in triggers:
    name = t.get('name', 'unnamed')
    last = t.get('last_fired_at', 'never')
    conns = [c.get('name','?') for c in t.get('mcp_connections', [])]
    print(f'{name}: last_fired={last}, connectors={conns}')
"

# Check latest Claude Code version (informational — fix is platform-side, not client):

npm view @anthropic-ai/claude-code version 2>/dev/null