agent-dispatch

A lightweight, platform-agnostic skill that acts as an agent registry and just-in-time router. Instead of pre-installing 130 agents (211k tokens at startup), your AI agent carries a compact keyword index (2k tokens) and downloads specialists on demand — mid-session, the moment it needs them.

Works with Claude Code, OpenClaw, Cursor, Codex, and any platform that supports the SKILL.md format.

The problem: pre-installing dozens of subagents bloats your context window. Not installing them means your agent does everything itself. Orchestrator agents are overkill — you'd spawn an agent just to pick an agent.

The solution: a single skill file with a TOML index mapping keywords to agent names and their download locations. When a specialist is needed, the skill teaches your agent to fetch it from GitHub, read its instructions, and dispatch — all in one seamless flow.

How it works

flowchart TD
    A["🧑‍💻 You're coding"] --> B["Agent hits a specialized task"]
    B --> C["🔍 Skill activates on keyword match"]

    C --> D["📖 Index lookup: security = security-auditor:quality"]
    D --> E{"🗂️ Cached locally?"}

    E -- "✅ YES" --> G["📄 Read agent instructions"]
    E -- "❌ NO" --> F["⬇️ Download from GitHub"]
    F --> G

    G --> H["🚀 Spawn Task with inline prompt"]
    H --> I["🔒 Work done in separate context window"]
    I --> J["📋 Results returned to main conversation"]
    J --> K["💾 Agent cached for next time"]

    style A fill:#0d1117,stroke:#58a6ff,color:#c9d1d9
    style B fill:#0d1117,stroke:#58a6ff,color:#c9d1d9
    style C fill:#0d1117,stroke:#3fb950,color:#3fb950,stroke-width:2px
    style D fill:#0d1117,stroke:#3fb950,color:#c9d1d9
    style E fill:#161b22,stroke:#f0883e,color:#f0883e,stroke-width:2px
    style F fill:#0d1117,stroke:#f778ba,color:#f778ba
    style G fill:#0d1117,stroke:#3fb950,color:#c9d1d9
    style H fill:#0d1117,stroke:#a371f7,color:#a371f7,stroke-width:2px
    style I fill:#0d1117,stroke:#a371f7,color:#c9d1d9
    style J fill:#0d1117,stroke:#58a6ff,color:#c9d1d9
    style K fill:#0d1117,stroke:#3fb950,color:#3fb950

Loading

Skills vs agents vs MCP

Mechanism	When loaded	Context impact	Best for
Skill	On-demand during conversation	Into main context (~minimal)	Routing, instructions, capabilities
Agent	When dispatched	Own context window (zero main impact)	Delegated specialized work
MCP	At startup before model loads	Consumes context upfront	External tool integrations

This skill bridges the gap: it's a skill (lightweight, on-demand) that routes to agents (isolated, specialized), downloading them just-in-time if they're not already installed.

Installation

Recommended: skill only (JIT mode)

Install just the dispatch skill — agents download automatically when needed:

git clone https://github.com/userFRM/agent-dispatch.git
cd agent-dispatch
./scripts/install.sh --skill-only

This installs only the dispatch skill (~2k tokens). Agents are fetched from GitHub on demand during your session and cached locally for future use.

OpenClaw (via ClawHub)

clawhub install agent-dispatch

Full install (skill + starter pack)

git clone --recurse-submodules https://github.com/userFRM/agent-dispatch.git
cd agent-dispatch
./scripts/install.sh

This installs the skill plus 10 pre-cached starter agents from the VoltAgent submodule.

Manual (any platform)

Copy skills/agent-dispatch/SKILL.md to your platform's skill directory:

Platform	Skill path
Claude Code	~/.claude/skills/agent-dispatch/SKILL.md
OpenClaw	~/.openclaw/skills/agent-dispatch/SKILL.md
Cursor	.cursor/skills/agent-dispatch/SKILL.md

Getting agents

With JIT mode (default), agents download automatically when needed. You can also pre-install them:

Fetch from GitHub by category

# List available categories
./scripts/fetch-agents.sh --list

# Fetch specific categories
./scripts/fetch-agents.sh --category quality
./scripts/fetch-agents.sh --category infra

# Fetch a single agent
./scripts/fetch-agents.sh --single debugger:quality

# Fetch all 130 agents from VoltAgent
./scripts/fetch-agents.sh --all

# Fetch from 0xfurai instead
./scripts/fetch-agents.sh --source 0xfurai

Available categories:

Category	Key	Agents	Examples
Core development	core	10	api-designer, frontend-developer
Language specialists	languages	26	typescript-pro, python-pro, rust-engineer
Infrastructure	infra	16	docker-expert, kubernetes-specialist
Quality and security	quality	14	code-reviewer, debugger, penetration-tester
Data and AI	data	12	data-scientist, llm-architect
Developer experience	devex	13	refactoring-specialist, documentation-engineer
Specialized domains	domains	12	fintech-engineer, game-developer
Business and product	business	11	product-manager, ux-researcher
Meta orchestration	meta	10	multi-agent-coordinator
Research and analysis	research	6	research-analyst, competitive-analyst

Git submodules (pinned versions)

git submodule update --init
cp vendors/voltagent/categories/04-quality-security/*.md ~/.claude/agents/

Usage

Automatic (recommended)

Once installed, Claude auto-consults the dispatch index when it encounters specialized tasks. If the agent isn't cached locally, it downloads it from GitHub, reads its instructions, and dispatches — all transparently. No action needed.

Manual

/agent-dispatch

Regenerate index from your agents

./scripts/generate-index.sh              # Print to stdout
./scripts/generate-index.sh --install    # Update SKILL.md in-place

Dispatch decision logic

The skill instructs Claude to dispatch when:

• The task is clearly specialized (security review, performance profiling)
• The agent can work independently with clear inputs/outputs
• The work would consume significant main context

And to do it inline when:

• The task is a quick one-liner
• Tight back-and-forth with the user is needed
• The user explicitly asks to handle it directly

Uninstall

# Remove skill only (agents untouched)
./scripts/uninstall.sh

# Remove skill + starter pack agents
./scripts/uninstall.sh --agents

# Remove skill + ALL agents in directory (with confirmation)
./scripts/uninstall.sh --all

Cross-platform usage

All scripts respect environment variables for platform-agnostic installs:

# Override agent directory (default: ~/.claude/agents)
AGENTS_DIR=~/.openclaw/agents ./scripts/fetch-agents.sh --all

# Override skill directory (default: ~/.claude/skills/agent-dispatch)
SKILL_DIR=~/.openclaw/skills/agent-dispatch ./scripts/install.sh

# Override SKILL.md path for index regeneration
SKILL_FILE=~/.openclaw/skills/agent-dispatch/SKILL.md ./scripts/generate-index.sh --install

Known limitations

• Each keyword maps to exactly one agent (TOML requires unique keys)
• Downloaded agents are cached permanently — delete the .md file to force re-download
• If you are offline, agents not already cached will be unavailable (the agent handles the task itself)
• ZIP downloads from GitHub don't include submodules — clone with --recurse-submodules or use the fetch script

Customization

Edit ~/.claude/skills/agent-dispatch/SKILL.md to:

• Add keywords for your specific workflows
• Remove agents you don't need
• Add categories for your domain

The index format is TOML with category suffixes:

keyword = "agent-name:category"

Works with

This skill is agent-source agnostic. It works with agents from:

• VoltAgent/awesome-claude-code-subagents (130 agents, MIT)
• 0xfurai/claude-code-subagents (100+ agents, MIT)
• affaan-m/everything-claude-code
• Your own custom agents in ~/.claude/agents/

Contributing

See CONTRIBUTING.md.

License

MIT

Acknowledgments

• Agent definitions from VoltAgent and 0xfurai, both MIT licensed
• Built on Claude Code's skill system