mlua-batteries 0.5.1

Batteries-included standard library modules for mlua
Documentation

mlua-batteries

Batteries-included standard library modules for mlua.

Lua 5.4 scripts gain access to JSON, environment variables, filesystem, HTTP, hashing, LLM chat completion, structured async tasks, and SQLite-backed storage — all behind a configurable policy layer that can sandbox untrusted code.

Core modules (json, env, path, time, fs, http, hash, llm, string, regex, validate, log, uuid, base64, schema, sandbox) are synchronous (blocking) and require no async runtime. The optional task module requires a tokio current-thread runtime driving a LocalSet (see the Async modules section), as do the std.sql / std.kv bridges in the companion mlua-batteries-sqlite crate.

Modules

Module Feature flag Description
json json JSON encode / decode (serde_json)
env env Environment variable access (overlay-safe set)
path path Path manipulation (pure computation + absolute)
time time Timestamps, sleep, measure
string string String utilities beyond Lua's built-ins
regex regex Regex match / replace (regex crate)
validate validate Lightweight value validation helpers
log log Bridge to the host's log facade
uuid uuid UUID v4 / v7 generation
base64 base64 Base64 encode / decode
fs fs File I/O, walk, glob, rename, symlink (Unix) (walkdir + globset)
http http HTTP client (ureq)
hash hash SHA-256 hashing (sha2)
llm llm Chat completion — OpenAI, Anthropic, Ollama
schema schema JSON Schema validation (schema-bridge)
sandbox sandbox Capability-based filesystem sandbox (cap-std)
proc proc Typed argv pipeline spawn — no shell involved
watch watch Filesystem versioning watcher — content-addressed store + JSONL journal (notify)
task task Structured async tasks with cooperative cancellation (requires tokio)

std.sql / std.kv moved to the mlua-batteries-sqlite crate in 0.5.0 — see SQLite below.

Default features: json, env, path, time, string, validate. Enable everything: full.

Quick start

[dependencies]
mlua-batteries = "0.3"
use mlua::prelude::*;

let lua = Lua::new();
mlua_batteries::register_all(&lua, "std").unwrap();

lua.load(r#"
    local data = std.json.decode('{"name":"Lua"}')
    print(data.name)                     -- Lua
    print(std.env.get("HOME"))           -- /Users/...
    print(std.path.join("/tmp", "a.txt"))-- /tmp/a.txt
    print(std.time.now())                -- 1740000000.123
"#).exec().unwrap();

Per-module registration

For integration with mlua-pkg or custom loaders:

for (name, factory) in mlua_batteries::module_entries() {
    // name: "json", "env", ...
    // factory: fn(&Lua) -> LuaResult<LuaTable>
    let table = factory(&lua).unwrap();
}

Sandboxing

The default configuration uses Unrestricted policies — Lua scripts can access any file, URL, and env var. For untrusted scripts, use Sandboxed (requires the sandbox feature):

[dependencies]
mlua-batteries = { version = "0.3", features = ["full"] }
use mlua::prelude::*;
use mlua_batteries::config::Config;
use mlua_batteries::policy::Sandboxed;

let lua = Lua::new();
let config = Config::builder()
    .path_policy(Sandboxed::new(["/app/data"]).unwrap().read_only())
    .max_walk_depth(50)
    .build()
    .expect("invalid config");

mlua_batteries::register_all_with(&lua, "std", config).unwrap();

Policy layers

Policy Controls Built-in options
PathPolicy Filesystem access Unrestricted, Sandboxed (cap-std)
HttpPolicy Outbound URLs Unrestricted
EnvPolicy Env var read/write Unrestricted
LlmPolicy LLM provider/model access Unrestricted

All policies are trait objects — implement the trait to create custom policies.

LLM module

Built-in providers: OpenAI, Anthropic, Ollama. API keys are read from environment variables (OPENAI_API_KEY, ANTHROPIC_API_KEY).

local resp = std.llm.chat({
    provider = "openai",
    model    = "gpt-4o",
    prompt   = "Hello!",
    system   = "You are helpful.",
    max_tokens  = 1024,
    temperature = 0.7,
})
print(resp.content)

-- Batch parallel requests
local results = std.llm.batch({
    { provider = "ollama",  model = "llama3.2", prompt = "Q1" },
    { provider = "openai",  model = "gpt-4o",   prompt = "Q2" },
})

Custom providers can be registered via mlua_batteries::llm::register_provider.

Async modules

task is async-first and requires a tokio current-thread runtime driving a LocalSet. It is not part of the default feature set — opt in explicitly.

[dependencies]
mlua-batteries = { version = "0.5", features = ["task"] }
tokio = { version = "1", features = ["rt", "macros"] }

task provides structured concurrency primitives (spawn, scope, with_timeout, sleep, checkpoint) with cooperative, level-triggered cancellation. with_timeout applies a 3-stage graceful-abort pattern (deadline → drain under grace_ms → hard-abort).

See the module-level rustdoc on src/task/mod.rs for the full API.

Async overrides for the blocking modules

The core modules stay synchronous, which means a script that calls std.time.sleep(5) or std.http.get(url) inside a task parks the single VM thread and every sibling task with it. async_overrides is the opt-in fix — call it after register_all + task::register and it replaces those entries in place:

mlua_batteries::register_all(&lua, "std")?;
mlua_batteries::task::register(&lua)?;
mlua_batteries::async_overrides::register_by_name(&lua, "std")?;

With it, std.time.sleep, std.proc.pipeline, std.http.* and std.fs.* no longer block the VM thread. Lua-side names, arguments, return values and error messages are unchanged — argument parsing, policy checks and result-table construction still happen on the VM thread, and only the bulk blocking work moves (tokio::time::sleep for time.sleep, tokio::task::spawn_blocking for the rest; policy resolution, including the max_read_bytes size stat, stays on the VM thread). It ships behind the same task feature and wants the same runtime: a current-thread runtime driving a LocalSet. Because the overrides are async functions, scripts using them run under call_async / eval_async, and a function passed to std.time.measure (which calls it synchronously) must not use an overridden entry.

std.time.sleep becomes cancel-aware in the same way std.task.sleep is. The spawn_blocking overrides do not interrupt work already handed to the blocking pool — cancelling the enclosing task cannot abort a running proc.pipeline, HTTP request or file read; each runs to completion (or, for a pipeline, to its own timeout_secs) while the VM thread stays free for siblings.

SQLite: std.sql / std.kv

The SQLite bridges live in companion crates. They were part of this crate up to 0.4.0 and moved out in 0.5.0. mlua-batteries-sqlite is the default: a host-owned rusqlite::Connection behind Arc<Mutex<_>>, with statements run in tokio::task::spawn_blocking. mlua-batteries-sqlite-isle is the variant for hosts that already run SQLite on a rusqlite-isle connection thread. The Lua-side API is the same; only the host wiring differs.

[dependencies]
mlua-batteries = { version = "0.5", features = ["task"] }
mlua-batteries-sqlite = "0.6"          # default, sync model
# ...or, on a rusqlite-isle AsyncIsle:
mlua-batteries-sqlite-isle = "0.5"

The reason for the split is the SQLite stack, not the code. libsqlite3-sys declares links = "sqlite3", so a build graph holds exactly one of its major versions, and cargo enforces that while resolving dependencies — meaning no crate can offer several clusters behind mutually exclusive features. Serving more than one cluster requires more than one published version line. Keeping that constraint on the small bridge crate leaves this crate's version line free for its own features, and leaves consumers who do not need SQLite free of a C library.

The companion crates' READMEs carry the cluster table and the wiring contract.

Configuration

Config::builder() exposes all tunable limits:

Setting Default Description
max_walk_depth 256 Max directory depth for fs.walk
max_walk_entries 10,000 Max entries from fs.walk / fs.glob
max_json_depth 128 Max nesting depth for JSON
http_timeout 30s Default HTTP request timeout
max_response_bytes 10 MiB Max HTTP response body size
max_sleep_secs 86,400 Max time.sleep duration
llm_default_timeout_secs 120 Default LLM request timeout
llm_max_response_bytes 10 MiB Max LLM response body size
llm_max_batch_concurrency 8 Max threads for llm.batch

Platform support

Unix only (Linux, macOS). Windows is not a supported target.

All path arguments are UTF-8. Non-UTF-8 Lua strings are rejected at the FromLua boundary.

MSRV

Rust 1.77 or later.

Contributing

Bug reports and feature requests are welcome — please open an issue. Pull requests are also appreciated.

License

Licensed under either of

at your option.