Agent Toolkit

AI agent tools. Ships with Navigator (MCP server) and Flow (Claude Code plugin).

Components

| Component | Pattern | Provides | | -------------------------------------------------------- | -------------------- | --------------------- | | Navigator | Finite State Machine | Workflow navigation | | Flow | Prompt injection | Orchestration ruleset |

Inversion of control: Traditional workflow engines use agents as tools. Navigator flips this - the AI orchestrator stays in control and uses the FSM for navigation.

Installation

# Install Flow plugin (includes Navigator MCP)
claude plugin marketplace add leclabs/agent-toolkit
claude plugin install flow@agent-toolkit

Quick Start (Flow)

/flow:task "add user authentication"
/flow:go

Architecture

flowchart TB
    subgraph "Claude Code Plugin"
        orchestrator["Orchestration Layer"]
    end

    subgraph Navigator["Navigator (MCP Server)"]
        sm["State Machine"]
        catalog["Workflow Catalog"]
    end

    subgraph "Your Project"
        workflows["Custom Workflows"]
        agents["Agent Definitions"]
    end

    orchestrator <-->|MCP Protocol| Navigator
    Navigator --> catalog
    Navigator --> workflows

Workflows

16 workflow templates in the catalog:

| Workflow | Purpose | | ------------------------- | --------------------------------------------------------- | | feature-development | Full lifecycle: plan, implement, test, review, PR | | bug-fix | Reproduce, investigate, fix, regression test | | bug-hunt | Parallel investigation: reproduce, archaeology, forensics | | quick-task | Minimal: understand, execute, verify | | agile-task | General: analyze, implement, test, review | | test-coverage | Analyze gaps, write tests, review | | refactor | Functional core / imperative shell restructuring | | ui-reconstruction | Extract semantic IR, rebuild UI, blind review | | context-optimization | Map connections, identify pathologies, improve | | build-review-murder-board | Build-review loop with 80% approval threshold | | build-review-quick | Build-review loop with basic sanity check | | batch-process | Dynamic fork/join for processing multiple items | | context-gather | Parallel context gathering | | execute | Single-step: just do the thing | | odd-even-test | Test workflow demonstrating fork with retry logic | | hitl-test | Minimal HITL recovery test |

License

ISC