Runbook Exec

PyPI version

Status: v0.2.2 — Beta. MCP server for Claude Desktop, markdown-fence fixes, Slack terminal fallback, Windows docs.

AI-driven CLI tool that transforms passive Markdown runbooks into executable automation with safety gates and audit trails.

At incident time (e.g., 3 AM), instead of a human manually following a runbook step-by-step, runbook-exec reads the Markdown runbook, classifies each step by risk level using Claude, executes safe steps autonomously, requests human approval for risky operations via Slack, and produces a tamper-evident audit log.

Quickstart

1. Install

pip install runbook-exec

Windows users

On Windows, always install inside a virtual environment to ensure the runbook-exec command is available in your PATH:

python -m venv venv
venv\Scripts\activate
pip install runbook-exec
runbook-exec --help

If you see 'runbook-exec' is not recognized as an internal or external command, you are not inside an active virtual environment. Run venv\Scripts\activate first.

Free Anthropic credits

New Anthropic accounts get $5 in free credits at console.anthropic.com — enough for hundreds of runbook-exec validate runs.

2. Set environment variables

Linux / macOS:

export ANTHROPIC_API_KEY=sk-ant-...          # Required
export SLACK_BOT_TOKEN=xoxb-...              # Required for approval workflow
export SLACK_APP_TOKEN=xapp-...              # Required for Socket Mode

Windows CMD:

set ANTHROPIC_API_KEY=sk-ant-...
set SLACK_BOT_TOKEN=xoxb-...
set SLACK_APP_TOKEN=xapp-...

Windows PowerShell:

$env:ANTHROPIC_API_KEY = "sk-ant-..."
$env:SLACK_BOT_TOKEN = "xoxb-..."
$env:SLACK_APP_TOKEN = "xapp-..."

These commands set variables for the current terminal session only. For permanent configuration on Windows, add them via Control Panel → System → Advanced → Environment Variables.

3. Run an example runbook

# Dry-run first to see what would happen
runbook-exec run examples/disk-full.md --dry-run

# Validate risk levels without executing
runbook-exec validate examples/disk-full.md

# Execute for real
runbook-exec run examples/disk-full.md

Platform support

Tested on Linux, macOS, and Windows. The shell executor uses subprocess and runs whatever shell command you provide — runbook commands need to be valid for the host OS. Most realistic runbooks target Linux. On Windows, use cmd.exe-compatible commands or wrap PowerShell via powershell -Command "...".

Runbook formatting

runbook-exec parses Markdown using the CommonMark spec. Two rules matter in practice:

Numbered lists only. Only numbered (1., 2., ...) list items become steps. Bullet points (-, *) are ignored.

Code blocks must be indented under their list item. A fenced code block is only associated with a step if it is indented at least as far as the list item's content. This is standard CommonMark behaviour — a code block at the left margin is not part of the list.

✅ Correct — code block indented under the list item:

1. Check disk usage

   ```bash
   df -h

❌ Wrong — code block at the left margin (will not be extracted as the step's command):
```markdown
1. Check disk usage

```bash
df -h

When in doubt, run `runbook-exec validate <runbook.md>` — steps with `command=None` in the output mean the code block was not picked up.

## Subcommands

### `run`

Execute a runbook end-to-end with safety gates and audit logging.

```bash
runbook-exec run <runbook.md> [OPTIONS]

Options:
  --dry-run              Simulate execution without running commands
  --incident-id TEXT     Identifier used in the audit log filename
  --auto-approve TEXT    Auto-approve steps at or below this risk level:
                         read_only | modifying | destructive
  --no-llm-context       Disable post-step LLM decision calls

validate

Parse and classify a runbook without executing any commands. Useful for CI/CD.

runbook-exec validate <runbook.md>

replay

Display a previous runbook execution from an audit log.

runbook-exec replay <audit-log.json>

Configuration

Create a .runbook-exec.toml file in your project root to set defaults:

llm_model = "claude-sonnet-4-5"
slack_channel = "#incidents"
timeout_seconds = 300
auto_approve_level = "read_only"   # "read_only" | "modifying" | "destructive"
audit_log_dir = "./runbook-exec-logs"

Environment variables

Variable	Required	Description
ANTHROPIC_API_KEY	Yes	Anthropic API key for Claude
SLACK_BOT_TOKEN	For approvals	Slack bot token (xoxb-...)
SLACK_APP_TOKEN	For approvals	Slack app token for Socket Mode (xapp-...)

Configuration precedence

CLI flags > .runbook-exec.toml > built-in defaults

Slack fallback

If Slack is not configured (missing SLACK_BOT_TOKEN or SLACK_APP_TOKEN), runbook-exec falls back to interactive terminal prompts for approvals and failure direction. A banner is shown at startup indicating which mode is active.

Slack is recommended for production use because terminal prompts require an active SSH or console session. Terminal mode is useful for local testing and development.

Risk levels

Each step is classified as one of three risk levels:

Level	Description	Examples
read_only	Observes system state only	df, ls, cat, kubectl get, ps
modifying	Changes state but recoverable	systemctl restart, logrotate, file edits
destructive	Deletes data or requires sudo	rm -rf, DROP TABLE, kubectl delete

By default, read_only steps run automatically. modifying and destructive steps require Slack approval.

Safety bias principle

When the classifier is uncertain between two risk levels, it always chooses the more cautious (higher risk) level. This is enforced in the LLM prompt, not in post-processing, so the reasoning reflects the actual decision.

LLM data disclosure

runbook-exec sends data to the Anthropic API in two scenarios:

1. Step classification (always): The step text and extracted command are sent to Claude to determine the risk level. This happens for every step before any execution begins.

2. Post-step decisions (unless --no-llm-context is set): After each step executes successfully, the step text, command, exit code, and stdout/stderr output are sent to Claude to decide whether to continue, skip subsequent steps, or abort.

What is NOT sent: Slack tokens, audit log contents, or any data from steps that have not yet executed.

--no-llm-context flag

Use --no-llm-context to prevent command output from being sent to the Anthropic API:

runbook-exec run runbook.md --no-llm-context

When this flag is set:

• Step classification still uses the LLM (only step text and command are sent)
• Post-step decision calls are disabled
• Execution continues to the next step after each command completes
• A warning is displayed at startup

Use this flag when running runbooks that produce sensitive output (credentials, PII, internal hostnames) that should not leave your network boundary.

Audit logs

Every execution produces a tamper-evident audit log in ./runbook-exec-logs/ (configurable via audit_log_dir).

Audit logs use NDJSON format with a SHA-256 hash chain. Each entry includes a hash of the previous entry, making tampering detectable.

To replay and verify an audit log:

runbook-exec replay runbook-exec-logs/disk-full-20240115T030000Z-a3f1.json

Audit log files are never committed to source control (.gitignore includes runbook-exec-logs/).

CI/CD integration

Use runbook-exec validate in PR checks to lint runbooks before they merge — catching broken formatting, missing commands, and classification failures before they reach an on-call engineer at 3 AM.

# .github/workflows/validate-runbooks.yml
name: Validate runbooks

on:
  pull_request:
    paths:
      - 'runbooks/**/*.md'
      - 'docs/runbooks/**/*.md'

jobs:
  validate:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4

      - uses: actions/setup-python@v5
        with:
          python-version: '3.11'

      - run: pip install runbook-exec

      - name: Validate all changed runbooks
        env:
          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
        run: |
          git diff --name-only origin/main...HEAD \
            | grep '\.md$' \
            | xargs -I{} runbook-exec validate {}

This only validates runbooks that were actually changed in the PR, not every runbook on every commit. The ANTHROPIC_API_KEY secret must be configured in your repository settings.

What validation catches:

• Markdown that fails to parse (broken formatting)
• Steps with no extracted command (command=None)
• LLM classification failures (malformed step text)
• Steps that would be silently skipped during a real execution

What validation does not do: execute any commands or connect to any infrastructure. It is safe to run on any CI machine.

Example runbook

See examples/disk-full.md for a realistic disk-full incident runbook with a mix of read_only, modifying, and destructive steps.

Use with Claude Desktop

runbook-exec can run as an MCP server, letting Claude Desktop execute runbooks through natural conversation — no terminal required.

Install and configure

Add to your Claude Desktop config file:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "runbook-exec": {
      "command": "runbook-exec",
      "args": ["mcp-server"],
      "env": {
        "ANTHROPIC_API_KEY": "sk-ant-...",
        "SLACK_BOT_TOKEN": "xoxb-...",
        "SLACK_APP_TOKEN": "xapp-..."
      }
    }
  }
}

Restart Claude Desktop. The runbook-exec tools will appear automatically.

Example conversations

"Validate the disk-full runbook at /runbooks/disk-full.md"

"Dry-run our deployment runbook and show me what would need approval"

"Show me what happened during last night's incident from the audit log"

"Run the disk-full runbook for real" ← Claude will pause at each risky step and ask for your approval before continuing

Safety

• execute_runbook defaults to dry_run=True — Claude cannot execute real commands unless you explicitly say "run for real" or "not dry-run"
• All approval gates apply: modifying and destructive steps always require your explicit approval, whether called from CLI or Claude
• Full audit log written for every execution, including MCP-initiated runs
• If Slack is configured, approvals go to Slack as normal
• If Slack is not configured, Claude handles approvals conversationally: it shows you the pending step, waits for "yes/go ahead", then resumes

Development

# Install in development mode
pip install -e ".[dev]"

# Run tests
pytest

# Run tests with coverage
pytest --cov=runbook_exec --cov-report=term-missing

# Lint
ruff check .

Install from source (development)

To contribute or run from a local clone:

git clone https://github.com/giridharpandurangi/runbook-exec.git
cd runbook-exec
pip install -e ".[dev]"