Claude Code

For headless Claude Code (claude -p from cron, CI, or any script) — not for interactive sessions, where Remote Control already covers you.

1. Add the MCP server

Fastest path — register the remote server directly:

claude mcp add --transport http pendnt https://api.pendnt.dev/mcp \
  --header "Authorization: Bearer $PENDNT_API_KEY"

Or drop a .mcp.json into your project so it's checked in and shared with anyone running the repo's headless jobs:

{
  "mcpServers": {
    "pendnt": {
      "type": "http",
      "url": "https://api.pendnt.dev/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_PENDNT_API_KEY"
      }
    }
  }
}

Don't commit the raw key. Reference an environment variable through your process manager/CI secrets, or generate .mcp.json at deploy time from a template.

2. Run headless with the MCP config

crontab -e
# 0 3 * * * cd /path/to/repo && claude -p "Deploy release/1.4.0 to prod" --mcp-config .mcp.json

Now any tool call the agent makes — including mcp__pendnt__request_approval — is available to it directly. A prompt like "before deploying, call request_approval with the diff summary" gets you the durable-approval flow with zero extra plumbing.

3. Wire it into Claude Code's own permission system

Calling the tool directly (step 2) works, but Claude Code also has a built-in permission system for every tool call the agent makes — not just the ones you explicitly prompt it to make. The pendnt Claude Code plugin ships a hook and a --permission-prompt-tool adapter so a permission decision that would otherwise need a TTY (or fall back to --dangerously-skip-permissions) routes through request_approval instead:

claude -p "Deploy release/1.4.0 to prod" \
  --mcp-config .mcp.json \
  --permission-prompt-tool mcp__pendnt__request_approval

With this flag, any permission prompt Claude Code would normally need an interactive terminal for is instead sent through pendnt to your channels — the agent blocks (or resumes later with the returned request_id) exactly like a direct request_approval call, but without you having to prompt for it explicitly in every task.

4. What the answer looks like

[telegram 03:14] "Deploy release/1.4.0 to prod?  [Approve] [Deny]"
[operator] taps Approve

< { "status": "approve", "by": "you", "at": "2026-08-29T03:15:02Z" }
$ agent resumes, runs the deploy, exits 0

Also see

• MCP tools reference — every tool, including notify, wait_for_event, and schedule_wakeup.
• Claude Agent SDK — for a long-running SDK daemon instead of one-shot claude -p runs.