slop-mcp

MCP orchestrator that aggregates multiple Model Context Protocol servers behind 8 meta-tools. Connect any number of MCPs without bloating your agent's context window.

Without slop-mcp:  50 MCPs x 20 tools = 1,000 tool definitions in context
With slop-mcp:     50 MCPs x 20 tools = 8 tool definitions in context

Install

uvx slop-mcp serve

Or install globally:

pip install slop-mcp

Also available via npm (npx @standardbeagle/slop-mcp), Go, and binary releases.

Configure

Add to Claude Desktop:

{
  "mcpServers": {
    "slop": {
      "command": "uvx",
      "args": ["slop-mcp", "serve"]
    }
  }
}

Add to Claude Code:

claude mcp add slop-mcp -- uvx slop-mcp serve

Windows users: Use cmd /c to wrap the command if path handling corrupts the args:

{
  "mcpServers": {
    "slop": {
      "command": "cmd",
      "args": ["/c", "uvx", "slop-mcp", "serve"]
    }
  }
}

Define MCP servers in .slop-mcp.kdl:

mcp "github" {
    transport "sse"
    url "https://mcp.github.com/sse"
}

mcp "jira" {
    transport "streamable"
    url "https://mcp.atlassian.com/v2/mcp"
}

mcp "lci" {
    command "lci" "mcp"
}

The 8 Meta-Tools

Examples

Cross-MCP orchestration with SLOP

Chain tools across multiple MCPs in a single run_slop call — intermediate results stay out of the agent's context:

# Create Jira tasks from unread emails matching a filter
emails = gmail.search_messages(query: "label:action-needed is:unread")
for email in emails {
    jira.create_issue(
        project: "OPS",
        summary: email["subject"],
        description: format("From: {}\n\n{}", email["from"], email["snippet"]),
        issue_type: "Task"
    )
    gmail.modify_message(id: email["id"], remove_labels: ["UNREAD"])
}
emit(created: len(emails))
# Agent sees only: {"created": 4}

# Index codebase structure into persistent memory for future sessions
results = lci.search(query: "public API endpoints")
endpoints = results
    | map(|r| {"path": r["file"], "name": r["symbol"], "kind": r["kind"]})
    | filter(|r| r["kind"] == "function")
mem_save("project", "api_endpoints", endpoints,
    description: "Public API endpoint inventory")
emit(indexed: len(endpoints))

# Generate a visual report from code analysis
stats = lci.search(query: "struct")
by_package = stats
    | map(|r| r["file"] | split("/") | first())
    | group_by(|pkg| pkg)
chart_data = by_package
    | items()
    | map(|pair| {"label": pair[0], "value": len(pair[1])})
    | sorted(|a, b| b["value"] - a["value"])
banana.create_chart(type: "bar", title: "Structs by Package", data: chart_data)

Features

• Progressive discovery — agents find tools via search_tools, not by loading everything upfront
• SLOP scripting — chain tool calls across MCPs and process results in a single run_slop call
• Lazy connections — MCP servers connect asynchronously with tool metadata caching
• Persistent memory — disk-backed mem_save/mem_load/mem_search across sessions
• Three-tier config — project-local > project > user config merging with KDL
• OAuth support — browser-based auth for MCPs like Figma, GitHub, Linear, Jira
• All transports — stdio, SSE, and streamable HTTP

Documentation

dev.standardbeagle.com/slop-mcp

License

MIT