Let Claude Code invoke itself.
Run agents on schedules, chat with them on Discord or Slack, and resume any session in your terminal. herdctl is an open-source orchestration layer that turns Claude Code into autonomous agents.
# Install herdctl globally
npm install -g herdctl
# Initialize your fleet
herdctl init
# Start your fleet
herdctl startWeb, Slack, or Discord — wherever your team already works
Chat with your agents through the built-in Web UI, Slack, or Discord. Send a message from your phone and pick up the conversation in your terminal. Every interaction is logged and every session is resumable.
Scale from one agent to an entire organization
Fleets can contain sub-fleets, so you can organize agents by team, project, or department. A marketing team runs analytics and SEO agents. A legal team runs compliance scanners. A DevOps team runs security audits and infrastructure monitors. Each team owns their fleet — and they all compose into one.
How herdctl works
Your fleet in a few YAML files
One herdctl.yaml defines your fleet. Each agent gets its own directory with a herdctl-agent.yaml. Secrets live in .env.
version: 1
fleet:
name: my-dev-fleet
description: Development agents
defaults:
model: sonnet
max_turns: 50
agents:
- path: ./agents/pr-reviewer
- path: ./agents/docs-writer
- path: ./agents/monitor
overrides:
docker:
enabled: trueRun agents in Docker containers
AI agents are inherently risky - they can be manipulated or confused into doing things they shouldn't. Docker isolation is the gold standard for containing that risk, and herdctl makes it dead simple.
name: my-agent
# Enable Docker isolation - that's it!
docker:
enabled: true
# Lock down what tools the agent can use
allowed_tools:
- Read
- Glob
- Grep
- Edit
- Write20-minute overview
Watch a complete walkthrough of herdctl's core features: fleet management, scheduled agents, chat integration, and session continuity.
Built-in Web Dashboard
See your entire fleet at a glance. Real-time status updates, live output streaming, and interactive chat - all from your browser. No external services required.
Key Features
Define agents that wake themselves up on schedules or triggers. Cron expressions, intervals, or one-off jobs - all from YAML config.
schedules:
daily_review:
type: cron
expression: "0 9 * * *"
prompt: |
Review all open PRs...If Claude Code can do it, your herdctl agent can do it. Same tools, same MCP servers. Lock down permissions per-agent.
allowed_tools:
- Read
- Edit
- WebSearch
- Bash(git)
- Bash(ssh)Connect agents to Discord or Slack. Message your agents from your phone and they continue working.
chat:
slack:
bot_token_env: SLACK_BOT_TOKEN
app_token_env: SLACK_APP_TOKEN
channels:
- "#dev-agents"Every job creates a real Claude SDK session. Resume that exact session in your terminal with claude --resume. Full context preserved.
# List recent sessions
herdctl sessions
# Resume in your terminal
claude --resume session-idCLI runtime uses your Claude Max subscription (much cheaper). SDK runtime uses API pricing*. Choose per-agent.
# Use Claude Max tokens
runtime: cli
# Or use Anthropic API
runtime: sdkTrigger actions on job completion. Post to Discord, run shell scripts, call webhooks - all configurable per-agent.
hooks:
after_run:
- type: discord
channel_id: "${CHANNEL}"
when: 'job.scheduleName'Agents come to you
Agents don't just wait to be checked on. Configure hooks to post results to Discord or Slack after every run. Get security findings, build failures, or daily summaries delivered straight to your team's channels.
A homelab agent reports security findings to Discord every morning
What can you build with herdctl?
Any task that's repetitive, time-sensitive, or benefits from continuous monitoring is a candidate for automation.
Review all open PRs every morning at 9am. Leave thoughtful comments, suggest improvements, flag security issues. Trigger via Slack or Discord: 'Hey, can you review PR #47?'
Point an agent at your codebase - it honors your CLAUDE.md, uses your slash commands, connects to your MCP servers. Chat with it on Slack or Discord from your phone.
Scan your codebase weekly for undocumented functions. Generate JSDoc comments, update README files, create examples. Review its own output before committing.
Daily scans for vulnerabilities, misconfigurations, and exposed secrets. Findings get reported to Discord. Run it long enough and it catches things humans miss.
Daily scans of competitor websites, pricing pages, and feature announcements. Get notified when something changes. Weekly summary reports to your team.
Monitor Reddit, Hacker News, Twitter for mentions of your product. Surface questions, complaints, and feature requests. Draft responses automatically.
Building herdctl in public
Follow along as I build and evolve herdctl. Technical deep-dives, design decisions, and lessons learned.
I justed added support for Composable Fleets to herdctl. As I build herdctl into more projects, I increasingly find myself creating a fleet of agents per project, and wanted a way to run them all from a central place so that I can juggle a bunch of things at once.
herdctl makes it really easy to run agentic workflows on a schedule. One of the first such workflows I added to herdctl itself was a daily security audit. This has been running for a few weeks now, so I wanted to share some early results.
herdctl now supports Slack and web connectors, enabling AI agents to interact with these platforms. The web connector gives your a
herdctl now supports running Claude Code Agents in Docker containers, significantly expanding your options for locking your agents down. It also added support for the claude code CLI runtime, ensuring your can use your inexpensive Claude Max Plan tokens.
herdctl is an orchestration layer for Claude Code. It lets your agents run on a schedule, as part of a fleet, and puts them right in your discord or slack channel.
Get started in under 5 minutes. Open source, free forever.
* The Claude Agents SDK does not officially support using your Claude Max account, but as of the time of writing it was nevertheless possible to set a CLAUDE_CODE_OAUTH_TOKEN environment variable and have the Agents SDK use your Max account. Long may it last.