MCP 服务器

agentforge

Provides 300+ MCP tools for orchestrating AI agents — including swarm coordination, self-learning memory, and task management — enabling MCP clients like Claude Code and Cursor to spawn, coordinate, and learn from specialist agents across sessions.

README

⚒️ AgentForge

Where AI agents are forged — orchestration for Claude Code & Codex

Maintained by Soulcynics404 · Harsshh

Quick Start · Install · Claude Code Integration · Swarm Commands · MCP Server · Docs

</div>

Agent = Model + Harness. The model writes; the harness gives it tools, memory, loops, sandboxes, and controls so it can actually work. AgentForge is that harness — the execution layer around Claude Code and Codex that adds 100+ specialized agents, coordinated swarms, self-learning memory, and enterprise guardrails. So agents don't just run, they collaborate.

AgentForge is an actively maintained fork of Ruflo (formerly claude-flow) by rUv, rebranded and maintained independently under the MIT license. All credit for the original architecture goes to the upstream project and its contributors.

✨ What you get

User ──▶ AgentForge (CLI / MCP) ──▶ Router ──▶ Swarm ──▶ Agents ──▶ Memory ──▶ LLM Providers
                ▲                                                        │
                └────────────────── Learning Loop ◀──────────────────────┘

	Feature	What it means
🧠	Self-learning memory	HNSW vector search across sessions — agents remember what worked and reuse it
🐝	Coordinated swarms	Hierarchical, mesh, and adaptive topologies with 100+ specialist agent types
🔀	Smart model routing	3-tier routing: deterministic codemod → Haiku → Sonnet/Opus (pay less, run faster)
🪝	Hooks + observability	Pre/post-edit hooks, session lifecycle tracking, 12 background workers
🔌	MCP server	300+ tools exposed to any MCP-capable client (Claude Code, Cursor, Zed…)
🧩	Plugin ecosystem	ADRs, DDD, security audits, cost tracking, browser automation & more
🛡️	Policy engine	Action approvals, decision ledger, safety envelopes for enterprise use

📦 Installation

Requirements: Node.js ≥ 20 · npm ≥ 9 · Claude Code or Codex CLI installed & authenticated (agents execute via these harnesses — AgentForge coordinates)

# Run directly (no install)
npx ruflo --version

# Or install globally
npm install -g ruflo

Note: the runtime package name is inherited from upstream (ruflo on npm). Project identity, branding, and maintenance are AgentForge's.

🚀 Quick Start

1️⃣ Initialize in your project

cd my-project
npx ruflo init

This creates:

my-project/
├── .claude/            # Claude Code settings + hook configs
├── .agentforge/        # memory DB, session state
├── .agents/skills/     # cross-agent skill registry
└── CLAUDE.md           # project rules wired for orchestration

2️⃣ Health check

npx ruflo doctor --fix      # verifies Node, MCP servers, memory DB, API keys

3️⃣ Launch Claude Code and just… talk

After init, open Claude Code normally — hooks auto-route tasks to swarms, learn from successful patterns, and coordinate agents in the background:

you: "Build a REST API with auth, tests, and docs"
     → planner agent breaks it down
     → coder agents build in parallel
     → reviewer agent checks output
     → memory stores what worked

🔗 Integrate with Claude Code

Automatic (recommended)

npx ruflo init wires everything: MCP registration, hooks in .claude/settings.json, skills, and CLAUDE.md.

Manual MCP registration

Add to your Claude Code MCP config (~/.claude.json or project .mcp.json):

{
  "mcpServers": {
    "agentforge": {
      "command": "npx",
      "args": ["ruflo", "mcp", "start"]
    }
  }
}

Then inside Claude Code you can call 300+ tools directly: memory_search, swarm_init, agent_spawn, task_orchestrate, neural_train, and more.

Install as a Claude Code plugin

# From the Claude Code prompt:
/plugin marketplace add Soulcynics404/AgentForge
/plugin install agentforge@AgentForge

The plugin bundles slash commands (/forge-init, /forge-swarm, /forge-status) plus the core skill set from plugin/skills/.

Integrate with other agents

Harness	How
Codex CLI	`npx ruflo init` detects Codex and writes `.agents/config.toml`
Cursor / Windsurf	Point their MCP config at the same `npx ruflo mcp start` server
Any MCP client	Same endpoint — AgentForge is harness-agnostic

🐝 Swarm coordination

# Initialize a swarm with a topology
npx ruflo swarm init --topology hierarchical --max-agents 8

# Spawn specialist agents
npx ruflo agent spawn --type coder    --name backend-dev
npx ruflo agent spawn --type reviewer --name qa-bot
npx ruflo agent spawn --type architect --name system-design

# Orchestrate a task across them
npx ruflo task create --description "Build auth module with tests"
npx ruflo swarm start

# Watch progress
npx ruflo swarm status

Topologies: hierarchical (queen-led) · mesh (peer-to-peer) · adaptive (self-organizing) · ring · star

<details> <summary><b>⚠️ Important: how execution actually works</b></summary>

AgentForge commands create coordination records — the actual building is done by your Claude Code / Codex agents. The correct mental model:

swarm init → sets up the coordination state
Your Claude Code session does the work (writes code, runs tests)
memory store → saves what worked for next time

Don't stop after calling a swarm command — keep working; the harness coordinates while you build. </details>

🧠 Memory system

# Store a pattern that worked
npx ruflo memory store --key "auth-pattern" --value "JWT+refresh rotation" --namespace patterns

# Search past learnings before starting new work
npx ruflo memory query "authentication best approach"

# Session persistence
npx ruflo session save && npx ruflo session resume --latest

Memory uses HNSW vector search over an SQLite/RVF store — semantically similar patterns surface even when keywords differ.

🖥️ MCP Server

Run AgentForge as a standalone MCP server any MCP client can consume:

npx ruflo mcp start              # stdio mode (default)
npx ruflo mcp start --port 3000  # HTTP/SSE mode
npx ruflo mcp tools              # list all exposed tools

Tool families: memory_*, swarm_*, agent_*, task_*, neural_*, hooks_*, policy_*, performance_* — full catalog.

📚 Documentation

Doc	Contents
User Guide	Complete reference: features, storage format, programmatic API
Plugin skills	Bundled skill library (agentdb, browser, neural, dual-mode…)
Architecture	Agent instructions & division-of-labor model
Changelog	Release history
Upstream docs	github.com/ruvnet/ruflo

🤝 Contributing

Issues and PRs welcome! For significant changes, open an issue first. Upstream contributions should go to ruvnet/ruflo; this fork tracks upstream and focuses on branding, packaging, and community maintenance.