mlua-swarm

A long-running swarm engine host that compiles flow.ir Blueprints and dispatches their agent steps to workers. A Blueprint declares a flow (step / seq / branch / loop / fanout / try / assign nodes) plus the agents it references; the engine resolves each agent to a backend (in-process Lua, a Rust function, a child process, or an interactive Operator) and drives the flow while recording task state.

Install

cargo install mlua-swarm-cli

This installs the mse binary with two subcommands:

mse serve   # HTTP + WS server (tasks, Blueprint store, WS Operator sessions)
mse mcp     # MCP adapter over stdio, for AI agents (Claude Code etc.)

One-shot Blueprint execution is available through the swarm_run MCP tool (exposed by mse mcp), which runs through the canonical TaskApplication.handle / TaskLaunchService::launch path.

MCP client config:

{ "command": "mse", "args": ["mcp"] }

Docker (no Rust toolchain required)

Also listed on the MCP Registry as io.github.ynishi/mlua-swarm:

{
  "mcpServers": {
    "mlua-swarm": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "ghcr.io/ynishi/mse:latest",
        "mcp"
      ]
    }
  }
}

Related MCP server: GRACE Orchestrator MCP

Documentation

Documentation is served from the code itself:

• API / architecture — rustdoc on docs.rs (the crate root doc is the architecture overview).

• Guides / samples / schema — bundled MCP resources under mse://, served by mse mcp and always version-matched to the binary: mse://guides/getting-started, mse://guides/bp-lifecycle (the develop → trial-run → operate workflow map), mse://guides/blueprint-authoring, mse://guides/mcp-tool-reference, mse://blueprints/samples/*, and the live Blueprint JSON Schema at mse://api/blueprint-schema (also available as the bp_schema tool).

Workspace crates

License

Licensed under either of Apache License 2.0 or MIT License at your option.