A Lua-based multi-agent orchestration runtime. Define complex multi-agent workflows as concise Lua scripts — the runtime handles scheduling, concurrency, checkpointing, and progress tracking automatically.

# Linux / macOS
curl -fsSL https://raw.githubusercontent.com/hi-youichi/luft/main/install.sh | sh

# Windows (PowerShell)
irm https://raw.githubusercontent.com/hi-youichi/luft/main/install.ps1 | iex

# Specific version
curl -fsSL https://raw.githubusercontent.com/hi-youichi/luft/main/install.sh | sh -s -- --version v0.4.2

# From source
cargo install luft-cli

Verify:

luft --version

Run an example workflow with the mock backend (no LLM required):

luft run --workflow examples/hello.lua --backend mock

Natural-language prompt — Luft generates a workflow plan via LLM, then executes it:

luft run "audit the codebase for security issues" -o report.md

Run a saved workflow with arguments:

luft run --workflow workflows/review_code.lua --args '{"target":"src/"}' --max-concurrency 4

You write a Lua orchestration script that spawns AI subagents to do the real work (reading files, writing code, web search, etc.). The script itself runs in a sandbox with no filesystem or shell access — it only holds the control flow, branching, and intermediate results.

┌──────────────────────────────────────────┐
│           User (CLI / Library / MCP)        │
├──────────────────────────────────────────┤
│         Lua Orchestration Runtime           │
│   agent · parallel · pipeline · phase      │
├──────────────────────────────────────────┤
│            Service Layer                    │
│   scheduling · checkpointing · events       │
├──────────────────────────────────────────┤
│            Backend Adapters                 │
│   OpenCode · Claude · Codex · Custom        │
└──────────────────────────────────────────┘

Key properties:

• Sandboxed scripts — no io, os, require, or shell access from Lua
• Checkpoint & resume — every run can be resumed from its last checkpoint
• Progress tracking — phases, agent counts, token usage, elapsed time
• Backend-agnostic — switch between AI providers without changing workflows
• MCP server — control workflows programmatically via JSON-RPC

--------------------------------------------
-- Goal:  Review source files in parallel
-- Arch:  files ==> parallel-review ==> report
-- Flow:  files[] -> results[] -> report
--------------------------------------------
meta = {
    reasoning = "Fan out file review across agents, collect findings",
    phases = {
        { label = "review", dynamic = true },
        { label = "report" },
    },
}

local FILES = { "src/main.rs", "src/lib.rs", "src/cli.rs" }

function main()
    phase("review", #FILES)

    local results = parallel(FILES, function(file)
        return agent({
            prompt = "Review " .. file .. " for security issues. "
                .. "Report any vulnerabilities found.",
        })
    end)

    phase("report")

    local findings = {}
    for i, r in ipairs(results) do
        if r.ok then
            table.insert(findings, { file = FILES[i], output = r.output })
        end
    end

    report({
        summary = "Reviewed " .. #FILES .. " files, "
            .. #findings .. " returned results",
        results = findings,
    })
end

More examples in examples/:

• hello.lua — simplest single-agent call
• parallel-demo.lua — parallel fan-out
• pipeline-demo.lua — streaming pipeline
• schema-demo.lua — structured output with schemas

use luft::Luft;

#[tokio::main]
async fn main() -> Result<(), luft::LuftError> {
    let luft = Luft::builder()
        .backend(MyBackend::new())
        .build()?;

    let outcome = luft.run_script(r#"
        function main()
            local result = agent({ prompt = "analyze code security" })
            report({ findings = result.output })
        end
    "#).await?;

    println!("{:?}", outcome.result);
    Ok(())
}

Luft includes a built-in MCP (Model Context Protocol) server. Any MCP-compatible agent can submit workflows, poll status, and read results:

luft mcp serve

Available MCP tools: workflow_execute, workflow_status, workflow_events, workflow_cancel, workflow_list_files, workflow_list_runs.

• Architecture overview
• SDK reference
• Library guide
• Tool reference
• Design documents
• Example workflows

MIT