Horde

One prompt. A horde of raiders. Bring back the goop.

Deploy prompts to multiple AI agents in parallel and collect independent opinions.

Inspired by counselors by Aaron Francis, but may go different paths. We'll see :) So, Horde is provided by Mathias Bachmann.

Install

Homebrew (recommended)

brew install codebeauty/tap/horde

Or as a cask:

brew install --cask codebeauty/tap/horde

Note: for now horde is macOS only (Apple Silicon and Intel).

Quick Start

horde wake

horde raid "review this authentication flow for security issues"

Commands

Command	Description
`horde raid [prompt]`	Deploy a prompt to AI agents in parallel
`horde summary latest`	Print the most recent run summary
`horde summary list`	List recent runs as detailed cards
`horde cleanup`	Remove old output directories
`horde wake`	Auto-discover installed AI CLIs and write config
`horde agents`	Manage configured agents (list, remove, test, discover, rename, add)
`horde loadouts`	Manage named loadouts of agents
`horde squads`	Manage raider squads (list, create, delete)
`horde stash`	Show resolved configuration
`horde doctor`	Check configuration and agent availability
`horde raiders`	Manage raiders (list, show, create, edit, reset)
`horde intel`	Print setup instructions for AI agent integration
`horde skill`	Print slash-command template for AI agent integration
`horde ls`	Alias for `horde agents list`

`horde raid [prompt]`

Deploy a prompt to AI agents in parallel.

# Send to all enabled agents
horde raid "review this authentication flow for security issues"

# Send to specific agents
horde raid -a claude-opus,gemini-3-pro "is this migration safe?"

# Use a named loadout
horde raid -f prompt.md --loadout smart

# Assign a raider to all agents
horde raid -R security "review this authentication flow"

# Pipe from stdin
echo "explain this error" | horde raid

# Gather context alongside the prompt
horde raid -c . "what does this diff do?"
horde raid -c src/core,src/adapters "review these modules"

Flags:
  -a, --agents <ids>       Comma-separated agent IDs
  -l, --loadout <name>     Named loadout from config
  -r, --read-only <mode>   enforced, bestEffort, or none
      --timeout <seconds>  Per-agent timeout (default: 540)
  -o, --output <dir>       Output directory override
      --json               Output manifest as JSON
  -f, --file <path>        Read prompt from file
      --dry-run            Show invocations without executing
  -c, --context <paths>    Gather context (comma-separated paths, or "." for git diff)
  -R, --raider <id>        Raider to apply to all agents (overrides per-agent config)
  -S, --squad <name>       Named squad of raiders (cross-product deploy)
      --yes                Skip confirmation prompts

--squad and --raider are mutually exclusive.

When running interactively with multiple agents and no --agents/--loadout flag, horde shows a numbered list for selection.

`horde summary`

View run summaries and browse run history.

horde summary latest              # Print the most recent summary
horde summary latest --path       # Print the run directory path (for scripting)
horde summary latest --json       # Print the run manifest (run.json) instead
horde summary list                # List recent runs as detailed cards
horde summary list --limit 5      # Show only the last 5 runs
horde summary list --json         # Output as JSON array of manifests

Both subcommands accept -o, --output-dir to override the output directory. Without it, the configured output directory is used (respecting project-level .horde.json overrides).

Example output:

--- 2026-02-23 00:36 ---
Prompt: Give me a compact idea how to build a SwiftUI app which use ...
Agents: claude (ok 22.171s), gemini (ok 38.357s)
Path:   agents/horde/give-me-a-compact-idea-...-1771803377

--- 2026-02-23 00:07 ---
Prompt: Audit my config for any potential issues
Agents: claude (ok 15.2s), codex (fail 28.1s), gemini (ok 22.4s)
Path:   agents/horde/audit-my-config-...-1771801594

`horde cleanup`

Remove old output directories.

horde cleanup                        # Remove runs older than 1 day (interactive)
horde cleanup --older-than 2w        # Remove runs older than 2 weeks
horde cleanup --dry-run              # Show what would be removed
horde cleanup --older-than 30m -y    # Remove runs older than 30 minutes, skip confirmation
horde cleanup --json                 # Output results as JSON

Supports duration suffixes: ms, s, m, h, d, w. A bare number is interpreted as days.

`horde wake`

Auto-discover installed AI CLIs and write config.

horde wake

Scans for claude, codex, gemini, and amp binaries in PATH, /opt/homebrew/bin, /usr/local/bin, and ~/.local/bin. Registers all model variants per adapter with recommended models enabled by default. Also syncs built-in raider presets to the raiders directory (existing customizations are preserved).

`horde agents`

Manage configured agents.

horde agents list              # Show all agents (alias: horde ls)
horde agents remove <id>       # Remove an agent
horde agents test [id...]      # Test agents with "Reply OK" prompt
horde agents discover          # Scan for new agents not yet configured

`horde loadouts`

Manage named loadouts of agents.

horde loadouts create smart --agents claude-opus,gemini-3-pro
horde loadouts list
horde loadouts delete smart

`horde raiders`

Manage raiders — role presets that shape how AI agents respond to the same prompt.

horde raiders list              # List all raiders (built-in + custom)
horde raiders show security     # Print raider contents
horde raiders create my-raider  # Create custom raider (opens $EDITOR)
horde raiders edit security     # Edit existing raider (opens $EDITOR)
horde raiders reset             # Re-sync built-in presets
horde raiders delete <id>       # Delete a raider (--force to ignore squad refs)

Horde ships 6 built-in raiders:

ID	Role
`security`	Security engineer — vulnerabilities, OWASP Top 10, attack surfaces
`performance`	Performance engineer — latency, memory, algorithmic complexity
`architect`	Software architect — SOLID, coupling, API design, extensibility
`reviewer`	Code reviewer — bugs, edge cases, readability, test gaps
`devil`	Devil's advocate — challenge assumptions, find flaws, argue the opposite
`product`	Product lead — user impact, acceptance criteria, prioritization

Raiders are markdown files stored in the raiders directory. Create custom raiders by adding .md files to that directory or using horde raiders create.

Creating a custom raider

# Option 1: Use the create command (opens $EDITOR)
horde raiders create golang-expert

# Option 2: Write the file directly
cat > ~/Library/Application\ Support/horde/raiders/golang-expert.md << 'EOF'
You are a senior Go developer reviewing for idiomatic patterns.

Focus on:
- Error handling (wrapping, sentinel errors, checking)
- Concurrency safety (goroutine leaks, race conditions, mutex usage)
- Interface design (small, consumer-defined)
- Naming conventions and package structure
- Performance (allocations, buffer reuse, unnecessary copies)

Be specific: point to the exact line, explain why it's wrong, and provide a corrected snippet.
EOF

The raider ID is the filename without .md — use letters, numbers, hyphens, underscores, and dots.

Using raiders

# Apply a raider to all agents for one run
horde raid -R security "review this authentication flow"

# Without -R, no raider is used — the prompt is sent as-is
horde raid "review this code"

Per-agent raiders in config

Assign default raiders to specific agents in config.json so each agent always adopts a different raider role:

{
  "tools": {
    "claude-opus": {
      "binary": "/usr/local/bin/claude",
      "adapter": "claude",
      "extraFlags": ["--model", "opus"],
      "enabled": true,
      "expert": "security"
    },
    "gemini-3-pro": {
      "binary": "/usr/local/bin/gemini",
      "adapter": "gemini",
      "extraFlags": ["--model", "gemini-2.5-pro"],
      "enabled": true,
      "expert": "architect"
    },
    "codex": {
      "binary": "/usr/local/bin/codex",
      "adapter": "codex",
      "enabled": true
    }
  }
}

In this setup, horde raid "review this code" sends the prompt to all three agents — Claude answers as a security engineer, Gemini as an architect, and Codex answers without a raider. The -R flag overrides all per-agent raiders for a single run.

When a raider is active, the prompt is prepended with the raider's role instructions. Each agent gets its own prompt file (<agent-id>.prompt.md) in the output directory. The original prompt.md is always preserved without raider modifications.

`horde squads`

Manage named squads of raiders. A squad is a reusable list of raider IDs. When used with --squad, horde creates a cross-product: each agent runs once per raider in the squad.

horde squads list                                     # List all squads
horde squads create code-review --raiders security,architect,reviewer
horde squads delete code-review

Using squads

# Run with a squad — creates agent x raider cross-product
horde raid -S code-review "review this module"

# Combine with agent selection (2 agents x 3 raiders = 6 runs)
horde raid -a claude-opus,gemini-3-pro -S code-review "review this code"

Each agent runs independently with each raider. The composite IDs use @ as separator (e.g., claude-opus@security, gemini-3-pro@architect). When the cross-product exceeds 8 runs, horde prompts for confirmation (skip with --yes).

`horde skill`

Print a 7-phase slash-command template for AI agent integration.

Configuration

Global config: ~/Library/Application Support/horde/config.json

Project overrides: .horde.json in the project root (merges with global; read-only mode can only be tightened, not loosened).

{
  "version": 1,
  "defaults": {
    "timeout": 540,
    "outputDir": "./agents/horde",
    "readOnly": "bestEffort",
    "maxParallel": 4
  },
  "tools": {
    "claude-opus": {
      "binary": "/usr/local/bin/claude",
      "adapter": "claude",
      "extraFlags": ["--model", "opus"],
      "enabled": true
    }
  },
  "groups": {
    "smart": ["claude-opus", "gemini-3-pro"]
  },
  "teams": {
    "code-review": ["security", "architect", "reviewer"]
  }
}

Field	Default	Description
`timeout`	540	Per-agent timeout in seconds
`outputDir`	`./agents/horde`	Base directory for run output
`readOnly`	`bestEffort`	`enforced`, `bestEffort`, or `none`
`maxParallel`	4	Max agents running concurrently

Per-agent fields:

Field	Description
`expert`	Default raider ID for this agent (overridden by `-R` flag)

Output Structure

Each run creates a timestamped directory:

agents/horde/
  review-auth-flow-1770676882/
    prompt.md              # Original prompt (without raider)
    run.json               # Manifest with metadata
    summary.md             # Heuristic summary (no LLM)
    claude-opus.md         # Claude's response
    claude-opus.stderr     # Claude's stderr
    claude-opus.prompt.md  # Per-agent prompt with raider (if raider used)
    gemini-3-pro.md        # Gemini's response
    gemini-3-pro.stderr    # Gemini's stderr
    claude-opus@security.md         # Squad run: Claude as security raider
    claude-opus@security.prompt.md  # Per-agent prompt with raider
    gemini-3-pro@architect.md       # Squad run: Gemini as architect
    gemini-3-pro@architect.stderr

When using --squad, output files use composite IDs (agent@raider).

Use horde summary latest to quickly view the most recent result, or --json on horde raid to get the manifest on stdout for programmatic consumption.

Duplicate Agent Runs

Run the same agent multiple times by repeating its ID:

horde raid -a claude-opus,claude-opus,claude-opus "review this code"

Creates three independent runs as claude-opus, claude-opus__2, claude-opus__3.

Development

See DEVELOPMENT.md for build instructions, adapter details, project structure, and security documentation.

Why horde?

Same prompt fan-out pattern as counselors, rewritten in Go.

	horde	counselors
Runtime deps	None	Node.js 20+
Install size	~3 MB	~102 MB
Startup	~4 ms	~120 ms
Peak memory	~5 MB	~65 MB

Both deploy to the same AI CLIs — response times depend on the AI provider, not the tool.

Name		Name	Last commit message	Last commit date
Latest commit History 47 Commits
.github/workflows		.github/workflows
assets		assets
cmd/horde		cmd/horde
docs		docs
internal		internal
.gitignore		.gitignore
.goreleaser.yaml		.goreleaser.yaml
DEVELOPMENT.md		DEVELOPMENT.md
LICENSE		LICENSE
Makefile		Makefile
README.md		README.md
go.mod		go.mod
go.sum		go.sum

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

Horde

Install

Homebrew (recommended)

Quick Start

Commands

`horde raid [prompt]`

`horde summary`

`horde cleanup`

`horde wake`

`horde agents`

`horde loadouts`

`horde raiders`

Creating a custom raider

Using raiders

Per-agent raiders in config

`horde squads`

Using squads

`horde skill`

Configuration

Output Structure

Duplicate Agent Runs

Development

Why horde?

About

Uh oh!

Releases 4

Contributors

Uh oh!

Languages

Folders and files

Latest commit

History

Repository files navigation

Horde

Install

Homebrew (recommended)

Quick Start

Commands

horde raid [prompt]

horde summary

horde cleanup

horde wake

horde agents

horde loadouts

horde raiders

Creating a custom raider

Using raiders

Per-agent raiders in config

horde squads

Using squads

horde skill

Configuration

Output Structure

Duplicate Agent Runs

Development

Why horde?

About

Topics

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases 4

Contributors

Uh oh!

Languages

`horde raid [prompt]`

`horde summary`

`horde cleanup`

`horde wake`

`horde agents`

`horde loadouts`

`horde raiders`

`horde squads`

`horde skill`