The embedded SQL database built for AI agents.

True MVCC · Built-in encryption · Per-session isolation · Rust & Python

FeatherDB is a full SQL database that compiles into your binary. No servers, no sockets, no C dependencies. It uses real database internals — B-trees, write-ahead logging, MVCC snapshot isolation — and adds the features AI agents actually need: concurrent writers, encryption at rest, API key auth, and per-conversation session isolation.

v1.0.0 "Louise" — First production release. True MVCC with write-write conflict detection, MCP server, all 22 TPC-H queries.

[dependencies]
featherdb = "1.0"

use featherdb::{Database, Result};

fn main() -> Result<()> {
    let db = Database::open("myapp.db")?;

    db.execute("CREATE TABLE tasks (
        id INTEGER PRIMARY KEY,
        title TEXT NOT NULL,
        status TEXT DEFAULT 'pending'
    )")?;

    db.execute("INSERT INTO tasks (id, title) VALUES (1, 'Build something cool')")?;

    let rows = db.query("SELECT * FROM tasks WHERE status = 'pending'")?;
    println!("Found {} pending tasks", rows.len());

    Ok(())
}

import featherdb

with featherdb.Database("myapp.db") as db:
    db.execute("CREATE TABLE tasks (id INTEGER PRIMARY KEY, title TEXT)")
    db.execute("INSERT INTO tasks VALUES (1, 'Build something cool')")

    for row in db.query("SELECT * FROM tasks"):
        print(f"Task #{row['id']}: {row['title']}")

JOINs (inner, left, right, full, cross) · Subqueries (IN, EXISTS, scalar, correlated) · CTEs · Window functions (ROW_NUMBER, RANK, DENSE_RANK) · UNION / UNION ALL · Views · GROUP BY + HAVING · CASE · 30+ built-in functions · Passes all 22 TPC-H queries

MVCC snapshot isolation · Multiple concurrent readers and writers · Deadlock detection · Configurable timeouts · Savepoints for partial rollback

B-tree engine · Write-ahead logging · Crash recovery · Adaptive compression (LZ4 / ZSTD) · Three eviction policies (Clock, LRU-2, LIRS)

AES-256-GCM page encryption with hardware acceleration · API key auth (Argon2id) · Table-level GRANT / REVOKE · Per-session isolation

Built-in Model Context Protocol server for AI agents. Tools: query, execute, list_tables, describe_table, create_table, get_status. Works with Claude Desktop, Cursor, and any MCP client.

# Start the MCP server
cargo run -p featherdb-mcp -- --db myapp.db

Cost-based optimizer · Index selection + join ordering · Predicate pushdown · Plan cache (5–10x on repeated queries) · Prepared statements

// Type-safe schema mapping
#[derive(Table)]
#[table(name = "users")]
struct User {
    #[column(primary_key)]
    id: i64,
    name: String,
    email: Option<String>,
}

// Fluent query builder
let users = QueryBuilder::select(&["name", "email"])
    .from("users")
    .where_gt("age", 21)
    .order_by("name", Ordering::Asc)
    .limit(10)
    .execute(&db)?;

┌──────────────────────────────────────────────────────────┐
│  PUBLIC API     Database · Transaction · QueryBuilder     │
├──────────────────────────────────────────────────────────┤
│  QUERY ENGINE   Parser → Planner → Optimizer → Executor   │
├──────────────────────────────────────────────────────────┤
│  MVCC           Snapshots · Version Chains · GC            │
├──────────────────────────────────────────────────────────┤
│  STORAGE        Buffer Pool · B-Tree · WAL · Encryption    │
└──────────────────────────────────────────────────────────┘

10 crates, clear boundaries:

See CONTRIBUTING.md for build instructions, test commands, and code style.

MSRV: Rust 1.70 · Platform: Linux (x86_64, ARM64) · macOS (Intel, Apple Silicon)

Dual-licensed under MIT or Apache 2.0, at your option.

v1.0.0 "Louise" — Each release is named after a Bob's Burgers or Futurama character.