Get started

Connect your agent

OpenFusion is a local MCP server, so it works with anything that speaks MCP — Claude Code, Cursor, Cline, Zed, Codex, Gemini CLI, and many more.

The easy way: run npx openfusion-setup. It detects your client, writes the right config, and installs the agent skill. The snippets below are for doing it by hand.

Using Seepient — our own agent?

It has a dedicated guide with setup and a full showcase walkthrough: OpenFusion + Seepient.

The server command

For an npm install, the server entry is always the same two lines — what differs is where each client wants to see them:

{
  "command": "npx",
  "args": ["-y", "openfusion-mcp"]
}

If you built from source, use the path to your build instead:

{
  "command": "node",
  "args": ["/absolute/path/to/OpenFusion/dist/index.js"]
}

You can also set OPENFUSION_HOME in the server's env to choose where config, keys, and the database live. Omit it to use the OS default — the active path prints on every startup.

Client support matrix

Client	Where it's configured	Auto-loads .mcp.json?	Skill?
Claude Code	.mcp.json or claude mcp add	✓	✓ .claude/skills/
ZCode	.mcp.json	✓	✓ .zcode/skills/
Seepient	~/.seepient/gateway.json or /gateway add	✗	✓ ~/.seepient/skills/
Cursor	.cursor/mcp.json	✗	rules .mdc
Cline / Roo Code	cline_mcp_settings.json	✗	.clinerules
Zed	context_servers in settings	✗	✗
Continue	~/.continue/config.yaml	✗	✗
Codex (OpenAI)	codex mcp add / ~/.codex/config.toml	✗	✗
Gemini CLI / Qwen / Kimi	~/.gemini/settings.json (and siblings)	✗	✗
Antigravity	mcp_config.json	✗	✓ .agent/skills/
opencode	opencode.json	✗	AGENTS.md
Hermes	~/.hermes/config.yaml	✗	✗
Claude Desktop	claude_desktop_config.json	✗	✗
Codebuff	.codebuff/mcp.json	✗	✓ .codebuff/skills/

No native MCP client?

Aider has no native MCP support yet (feature requested). Pi and OpenClaw connect through community adapters — follow the adapter's own README.

Per-client setup

Claude Code

The simplest path: open your project in Claude Code and approve the .mcp.json it auto-loads. To register globally instead:

claude mcp add openfusion -s user -- npx -y openfusion-mcp

ZCode

Add to ~/.zcode/cli/config.json under mcp.servers:

{
  "mcp": {
    "servers": {
      "openfusion": {
        "command": "npx",
        "args": ["-y", "openfusion-mcp"]
      }
    }
  }
}

Cursor

Create .cursor/mcp.json in your project (or ~/.cursor/mcp.json for all projects):

{
  "mcpServers": {
    "openfusion": {
      "command": "npx",
      "args": ["-y", "openfusion-mcp"]
    }
  }
}

Cline / Roo Code (VS Code)

Open the Cline sidebar → MCP servers → Edit MCP settings, which opens cline_mcp_settings.json, and add:

{
  "mcpServers": {
    "openfusion": {
      "command": "npx",
      "args": ["-y", "openfusion-mcp"],
      "disabled": false,
      "autoApprove": []
    }
  }
}

Zed

Add to ~/.config/zed/settings.json. Note the key is context_servers and command is an object:

{
  "context_servers": {
    "openfusion": {
      "command": {
        "path": "npx",
        "args": ["-y", "openfusion-mcp"]
      }
    }
  }
}

Seepient

Seepient hosts MCP servers through its gateway (enabled by default). Add OpenFusion to the targets array in ~/.seepient/gateway.json — the full walkthrough, including the showcase use-case, is in OpenFusion + Seepient:

{
  "targets": [
    {
      "kind": "mcp",
      "transport": "stdio",
      "command": "npx",
      "args": ["-y", "openfusion-mcp"],
      "description": "Fusion panel — many models, one judged answer",
      "tags": ["ai", "research"],
      "enabled": true
    }
  ]
}

Restart Seepient, verify with /gateway list, and install the skill into ~/.seepient/skills/openfusion/.

Continue

Edit ~/.continue/config.yaml:

mcpServers:
  openfusion:
    command: npx
    args:
      - -y
      - openfusion-mcp

Codex (OpenAI)

codex mcp add openfusion -- npx -y openfusion-mcp

Or edit ~/.codex/config.toml — note it's mcp_servers with an underscore:

[mcp_servers.openfusion]
command = "npx"
args = ["-y", "openfusion-mcp"]

Gemini CLI · Qwen Code · Kimi Code

All three are Gemini-CLI-family — same shape, different files (~/.gemini/settings.json, ~/.qwen/settings.json, ~/.kimi/settings.json):

{
  "mcpServers": {
    "openfusion": {
      "command": "npx",
      "args": ["-y", "openfusion-mcp"]
    }
  }
}

Antigravity

Create mcp_config.json in the project root (or .antigravity/mcp_config.json):

{
  "mcpServers": {
    "openfusion": {
      "command": "npx",
      "args": ["-y", "openfusion-mcp"]
    }
  }
}

opencode

Create opencode.json in the project root. Note the distinct shape — type, command as an array, environment:

{
  "mcp": {
    "openfusion": {
      "type": "local",
      "command": ["npx", "-y", "openfusion-mcp"]
    }
  }
}

Hermes

Edit ~/.hermes/config.yaml:

mcpServers:
  openfusion:
    command: npx
    args:
      - -y
      - openfusion-mcp

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS):

{
  "mcpServers": {
    "openfusion": {
      "command": "npx",
      "args": ["-y", "openfusion-mcp"]
    }
  }
}

Codebuff

Create .codebuff/mcp.json in the project root:

{
  "mcpServers": {
    "openfusion": {
      "command": "npx",
      "args": ["-y", "openfusion-mcp"]
    }
  }
}

Install the agent skill (recommended)

The setup command installs the openfusion skill automatically where the client supports it. The skill teaches your agent the working method: do the legwork first — read, search, reproduce — then bring a prepared dossier and call fusion once. Without it, agents tend to call fusion for everything, which wastes time and money.

If you'd rather install it by hand, copy the skill/ folder into your client's skills directory so it ends up as openfusion/SKILL.md plus openfusion/references/:

Client	Skills path
Claude Code	.claude/skills/openfusion/
ZCode	.zcode/skills/openfusion/
Antigravity	.agent/skills/openfusion/
Codebuff	.codebuff/skills/openfusion/

Let your agent do it

If your client can edit files and run commands, paste this and it will wire everything up:

Add the OpenFusion MCP server to your config. The server command is npx -y openfusion-mcp. Register it as openfusion in whatever MCP config this client reads, then tell me to restart so it loads. After that, confirm the fusion and open_dashboard tools are available.

Verify

After restarting your client:

fusion({ prompt: "In one sentence, why fuse multiple LLMs?" })

A consolidated answer means you're done. An error pointing at http://localhost:9077 means the panel isn't configured yet — finish the dashboard setup in Installation.

---

Next up

• Configuration — pick your panel and your judge
• Using the fusion tool — parameters, personas, async results