A command-line tool for running multiple Claude Code commands sequentially with automatic retry, rate limit handling, and output verification.

automate-claude wraps the Claude CLI tool to enable automated, unattended execution of multiple commands. It handles common automation challenges like:

• Sequential command execution - Run multiple Claude commands one after another
• Output verification - Uses Claude to verify each command completed successfully
• Automatic retry - If a command fails, automatically attempts recovery
• Rate limit handling - Detects rate limits and waits for reset before continuing
• Live streaming - Real-time output streaming with JSON parsing
• Detailed logging - Saves all output to timestamped log files

Pre-built static binaries are available on the releases page. Currently only tested on Ubuntu.

./automate-claude --help

Requires the Jai programming language compiler.

The provided binary was compiled with Jai version beta 0.2.018 (built on 11 October 2025).

jai automate-claude.jai

This produces the automate-claude executable.

mkdir -p output

# Build the image
docker build -t automate-claude-builder .

# Run the build (mount Jai compiler and get output)
docker run -v /path/to/jai:/jai -v $(pwd)/output:/output automate-claude-builder

# Run a single command
./automate-claude "refactor the utils module"

# Run multiple comma-separated commands
./automate-claude "fix lint errors","run tests","update documentation"

# Use slash commands
./automate-claude "/implement feature","write tests","fix lint"

Commands are executed sequentially. If any command fails, execution stops.

Sets the maximum time (in minutes) each command is allowed to run before being killed.

# Allow up to 2 hours per command for long-running tasks
./automate-claude --timeout 120 "refactor entire codebase"

# Quick timeout for simple commands
./automate-claude --timeout 5 "fix typo in README"

Note: At the moment, if any command is killed via timeout the entire program will exit.

Enables real-time output streaming. When enabled:

• Output is displayed immediately as Claude generates it
• Uses Claude's --output-format stream-json for proper streaming
• Output is simultaneously written to log files
• Useful for monitoring long-running tasks

./automate-claude --live "implement new feature"

Without --live, output is captured silently and only displayed in the final log files.

Adds the --dangerously-skip-permissions flag to Claude commands. This allows Claude to:

• Execute commands without confirmation prompts
• Write files without asking
• Run potentially destructive operations

Warning: Only use this in controlled environments where you trust the commands being executed.

./automate-claude --skip-perms "format all code files"

Note: This flag does not work when running as root. Use --headless instead.

Full automation mode designed for unattended execution, particularly in Docker containers or CI/CD pipelines or in WSL. When enabled:

• Sets IS_SANDBOX=1 environment variable
• Automatically enables --dangerously-skip-permissions
• Works correctly when running as root

# Run in a Docker container
./automate-claude --headless "run all tests","fix any failures"

# In a CI/CD pipeline
./automate-claude --headless --timeout 30 "lint code","run tests"

Repeats all commands multiple times. Useful for iterative tasks or ensuring consistency.

# Run the same workflow 3 times
./automate-claude --count 3 "check for issues","fix issues"

# This runs: check -> fix -> check -> fix -> check -> fix

The total number of runs is number_of_commands * count.

All output is saved to claude_runs/<timestamp>/:

claude_runs/
  2024-01-15_14-30-45/
    run_1.txt          # Output from first command
    run_1_check.txt    # AI verification log for first command
    run_2.txt          # Output from second command
    run_2_check.txt    # AI verification log
    run_2_retry.txt    # Retry output (if retry was needed)
    summary.txt        # Final summary of all runs

1. Parse commands - Split comma-separated command string
2. Create output directory - Timestamped folder under claude_runs/
3. For each command:
   • Run Claude with the command
   • Save output to log file
   • Use Claude to verify success/failure
   • If failed: attempt automatic retry
   • If retry fails: stop execution
4. Write summary - Record total time and success status

After each command, the tool runs a separate Claude instance to analyze the output and determine if the command succeeded. This catches cases where:

• Claude encounters errors but exits with code 0
• Work was partially completed
• Unexpected issues occurred

The verification prompt looks for:

• Error messages and exceptions
• Completion indicators
• Build errors or warnings

If AI verification fails, the tool attempts automatic recovery:

1. Creates a recovery prompt asking Claude to continue from where it left off
2. Points Claude to the previous run's log file for context
3. Runs the recovery command
4. Verifies the retry output

The tool detects rate limit messages in the format:

5-hour limit reached · resets 1pm (Australia/Brisbane)

When detected:

• Calculates wait time until reset
• Prints countdown status every 5 minutes
• Automatically resumes execution after the wait
• Adds a 2-minute buffer after reset time

# Single task
./automate-claude "fix all TypeScript errors"

# Multi-step workflow
./automate-claude "analyze codebase","create implementation plan","implement changes"

# Run in headless mode with timeout
./automate-claude --headless --timeout 30 \
  "run linter","fix lint errors","run tests"

# Stream output while running complex refactoring
./automate-claude --live --timeout 120 \
  "refactor authentication module","update all tests","verify no regressions"

# Keep improving until no more issues (run 5 iterations)
./automate-claude --count 5 --live \
  "find code quality issues","fix one issue"

# In Dockerfile or docker run
docker run -v $(pwd):/workspace myimage \
  ./automate-claude --headless "run tests"

• 0 - All commands completed successfully
• Non-zero - A command failed, timed out, or verification failed

Ensure claude is installed and in your PATH:

which claude
claude --version

Increase the timeout:

./automate-claude --timeout 120 "long running task"

Use --headless when running as root, or --skip-perms otherwise:

./automate-claude --headless "task requiring permissions"

The tool automatically handles rate limits. You can also:

• Use a different API key
• Check your usage limits

• Claude CLI (claude) installed and configured
• Jai compiler (for building from source)
• Linux/POSIX environment (uses script command for PTY)

See LICENSE file for details.