Agent Dispatcher: pricing, features, alternatives

Agent Dispatcher is a capability pack for Claude Code and Codex. It routes a request to one specialist role, loads a small set of skills and tools, and states what evidence will count as done. The repo nahid-sparktales/agent-dispatcher is Python, created 2026-09-19, MIT. On 2026-09-20 GitHub listed 16 stars and 0 forks. 16 stars is a heat signal, not traffic. The README counts 27 roles, 79 local skills, 31 external skills, 8 recipes, 19 MCP servers, and 50 detection signals. Those counts are first-party labels in the README; they are not a measured quality score.

It is instructions the host agent adopts in a session. Installing the pack does not start a team of agents or connect MCP servers. External skills and MCP entries are catalog references. Compare Agent Skills for a broader engineering pack, unlazy if you only needed acceptance gates, or Jev if you wanted a typed decision model. The README's optional Jev path stayed optional after a 162-case routing check on this repo: keyword routing was weak, Jev was better, Claude-style routing was best on that table. Treat the table as this author's run, not a public benchmark.

Key Features

Role owns the outcome: debugger, reviewer, UI designer, and others. A skill is a method. An MCP is a capability. A recipe is a workflow across roles. The README's design rule is that skills should stay narrower than an entire engineer role.
Context plan: the dispatcher names skills, project files, tools, guidance, and verification before the work starts. Compact output is one line. Verbose output lists dropped files and fallbacks.
Host installers: Codex uses python3 install_codex.py into ~/.agents/skills/agent-dispatcher/. Claude Code plugin: claude plugin marketplace add nahid-sparktales/agent-dispatcher then claude plugin install agent-dispatcher@agent-dispatcher. Manual Claude path: ./install.sh. Use one method per host. Native Windows is not covered; the README points at WSL or another Unix environment.
Direct role commands: /agent-reviewer, /agent-debugger, /agent-uidesigner, /agent-inventory, /agent-doctor, /agent-context, /agent-decision. Plugin installs prefix commands with agent-dispatcher:.
Python 3.10+, no third-party Python packages for installers and validation. CI covers Python 3.10-3.14 on Linux and Python 3.14 on macOS.

Limitation: routing to a role does not grant permission. Authorization still comes from you and the host. The pack is not a sandbox. Automatic session activation is off by default in Codex and needs hook trust review.

Use Cases

Large skill libraries where loading everything into every task adds conflicting guidance.
Reviews and debugging that should name the checks that actually ran.
Skip if the job is a one-line edit. The README says trivial tasks need no role or context plan.

Pricing

Piece	Price	Notes from first-party pages 2026-09-20
Pack	$0	MIT. Clone and install. About 16 stars.
Host / models	Your plan	Claude Code or Codex usage still bills as usual.
External MCP / skills	Whatever those vendors charge	Catalog only; the pack does not install them.

Getting Started

Clone nahid-sparktales/agent-dispatcher.
Codex: python3 install_codex.py. Claude plugin: marketplace add, then plugin install. Or ./install.sh for manual Claude.
New session. Codex: $agent-dispatcher Find why the tests are failing, fix the cause, and verify the fix. Claude plugin: /agent-dispatcher:agent-dispatcher with the same ask.
/agent-dispatcher status (or the plugin-prefixed form) to inspect activation.

First-party start: the README quick start and Claude Code plugins.

Frequently Asked Questions

Does it run extra agents in the background?

No. Roles are working instructions the current coding agent adopts. It can chain roles or assign them to subagents only when the host already allows that.

Should I turn on Jev routing?

Only if you want it. The README kept Claude/Codex-style routing as the default after its own 162-case table.

Can I force a role?

Yes. /agent-reviewer or $agent-dispatcher reviewer stays until you pick another or stop the dispatcher.

Alternatives

Agent Skills: a portable engineering skill set, not a role router.
unlazy: completion gates, not specialist routing.
Claude Code: the usual host for the plugin install.

Tips

Install only one method per host so commands and hooks do not duplicate.
Read /agent-context explain before you trust a surprising route.
Treat verification lines as a report of what ran, not a proof of correctness.

Conclusion

Agent Dispatcher is a new, small MIT pack for people whose skill folders got too large. Route the task, load less, and ask for evidence. Skip it for tiny edits, and do not confuse a role switch with extra permission.