Dragonfly

Opus-powered MCP plugin for Claude Code: 74 tools across 13 modules covering code intelligence, semantic search, persistent memory, and workflow orchestration.

Installation

npm install --save-dev @dragonflymcp/plugin

The postinstall script runs automatically and deploys all agents, skills, commands, hooks, and .mcp.json into your project. Restart Claude Code to activate.

To re-run setup manually at any time:

npx dragonfly-init

Setup is idempotent — existing files are never overwritten.

Modules

| Module | Tools | Description | |--------|-------|-------------| | AST | 7 | Code intelligence through AST analysis — call graphs, symbol lookup, reference finding | | Semantic | 3 | Semantic search with LRU-cached embeddings and incremental re-embedding | | Memory | 4 | Persistent semantic memory with episodic/semantic/procedural categorization | | Framework | 8 | Workflow orchestration with memory context, analytics snapshots, and bridge export | | State | 8 | Checkpoints, stories, provenance events, workflow session persistence | | Evolve | 4 | Genetic algorithm prompt optimization with real-failure training data | | Spec | 6 | Formal specification DSL for type-safe code generation | | Testing | 6 | Test generation, execution, coverage analysis, untested file discovery | | Repair | 5 | Self-repair, iterative refinement, test-driven repair with evolve test case capture | | Knowledge Graph | 7 | Entity/relation management with hybrid search (semantic + keyword + graph) | | Analytics | 5 | Cost/latency/quality benchmarks, pattern learning with evolve hints, drift detection | | Pipeline | 2 | WYSIWID composition DSL and execution planning | | Bridge | 4 | Cross-project memory federation with auto-export on workflow completion |

Integration Architecture

As of v1.3.x, modules are wired together into a connected system. Key integration points:

• Memory context on workflow start — dragonfly_start_workflow recalls similar past workflows and returns memory_context with outcomes and similarity scores.
• Testing guidance injection — Quality steps receive enriched prompts listing run_tests, run_tests_with_repair, analyze_coverage, and find_untested_files.
• Repair guidance on failure — Failed steps return repair_guidance pointing to self_debug, run_tests_with_repair, and iterative_refine.
• Analytics snapshot on completion — Completed workflows return analytics_summary with cost, latency, and quality metrics.
• Bridge auto-export on completion — High-confidence memories are automatically exported to the global bridge store.
• EmbeddingCache and EmbeddingRefresher — embed_project uses hash-based change detection and 4-parallel embedding with LRU caching.
• Spec in feature DSL — Medium and large feature workflows include the spec concept in their pipeline templates.
• Evolve hints from analytics — dragonfly_learn_patterns returns evolve_hint when high-confidence patterns accumulate.

See docs/integration.md for the full integration map, data flows, and response field reference.

The Evolve-from-Repair Loop

Repair events are automatically captured as evolve-test-case memories with no user action required. Over time, these accumulate real failure patterns from the project. When dragonfly_learn_patterns detects enough high-confidence patterns, it returns an evolve_hint suggesting prompt optimization. Running evolve_start with use_memory_test_cases: true loads these real failures as training data, producing project-specific prompt variants that Claude evaluates across generations. The result is skills that reflect actual project failure patterns rather than generic best practices.

See docs/integration.md for the full feedback loop diagram and design rationale.

What Gets Deployed

| Destination | Count | Source | |-------------|-------|--------| | .claude/agents/ | 18 agents | templates/agents/*.md | | .claude/skills/ | 43 skills | templates/skills/*.md.template | | .claude/commands/ | 31 commands | templates/commands/*.md.template | | .claude/hooks/ | 15 hooks + lib | templates/hooks/*.template | | .mcp.json | MCP server config | written by init script |

Documentation

| File | Description | |------|-------------| | docs/ast.md | AST analysis tools — call graphs, symbols, references | | docs/semantic.md | Semantic search, embedding cache, incremental re-embedding | | docs/memory.md | Persistent memory — store, recall, forget, evolve | | docs/framework.md | Workflow orchestration — start, advance, plan, compose | | docs/state.md | Checkpoints, stories, provenance, sessions | | docs/spec.md | Specification DSL — generate, save, import, export | | docs/testing.md | Test generation, execution, coverage, untested files | | docs/repair.md | Self-debug, iterative refine, test-driven repair | | docs/evolve.md | Genetic prompt optimization — start, submit, best, status | | docs/analytics.md | Benchmarks, pattern learning, drift detection | | docs/bridge.md | Cross-project memory export, import, search | | docs/integration.md | Full integration architecture and data flow reference |

Feature Flags

Modules are opt-in via environment variables. Set in your .mcp.json env block or shell:

DRAGONFLY_MEMORY_ENABLED=true
DRAGONFLY_FRAMEWORK_ENABLED=true
DRAGONFLY_STATE_ENABLED=true
DRAGONFLY_EVOLVE_ENABLED=true
DRAGONFLY_SPEC_ENABLED=true
DRAGONFLY_REPAIR_ENABLED=true
DRAGONFLY_KG_ENABLED=true
DRAGONFLY_ANALYTICS_ENABLED=true
DRAGONFLY_PIPELINE_ENABLED=true
DRAGONFLY_BRIDGE_ENABLED=true

AST, Semantic, and Testing modules are always enabled.

Environment Variables

PROJECT_ROOT                      # working directory (default: cwd)
DRAGONFLY_STATE_DB_PATH           # override state DB path
DRAGONFLY_MEMORY_DB_PATH          # override memory DB path
DRAGONFLY_INDEX_PATH              # override AST index path
DRAGONFLY_FRAMEWORK_CONTENT_ROOT  # override templates path
DRAGONFLY_DEBUG=true              # enable debug logging

Data Directory

Dragonfly stores all runtime state in data/ at the project root:

data/
├── index/       # AST index (rebuilt via index_project tool)
├── state.db     # checkpoints, stories, workflows, specs
└── memory.db    # memories, embeddings, knowledge graph

Add data/ to .gitignore. It is created automatically on first run.

MCP Configuration

The init script writes this to .mcp.json:

{
  "mcpServers": {
    "dragonfly": {
      "type": "stdio",
      "command": "node",
      "args": [
        "--no-wasm-tier-up",
        "--liftoff-only",
        "node_modules/@dragonflymcp/plugin/dist/index.js"
      ]
    }
  }
}

The WASM flags are required for tree-sitter stability.

Building from Source

git clone https://github.com/mkolb22/dragonfly-plugin
cd dragonfly-plugin
npm install
npm run build    # tsc — must be 0 errors
npm run test:run # vitest run

License

MIT AND Commons Clause — free to use, modify, and distribute. Commercial use requires licensor approval. Copyright 2026 Michael Kolb.