A CLI tool to bootstrap and enhance LLM agent-driven database learning. A TUI visualizes multi-transaction scenarios executed on real databases, letting you observe transaction anomalies and gotchas firsthand.

• Why ACID?
• Demo
• Features
• AI-guided learning
• Quick Start
• Controls
• Contributing
• License

1. Zero to hero — deep understanding of database behavior

Start from scratch and build real intuition for how databases handle concurrency. An AI coach walks you through each anomaly — dirty reads, lost updates, phantom reads, deadlocks — with live scenarios on a real database, not toy diagrams.

"I'm new to databases. Teach me ACID properties and transaction isolation from the ground up."

2. Interview prep and topic revision

Quickly revisit specific concepts before a system design interview or when a topic feels fuzzy. Predict outcomes, get corrected, understand why — the prediction-first loop makes knowledge stick.

"I have a system design interview tomorrow. Run me through isolation levels and the most common concurrency gotchas."

3. Testing database behavior during feature design

Unsure how your database will behave under concurrent writes? Write a scenario for your exact table structure and query pattern, run it, and see the real result before committing to an approach.

"I'm designing a booking system. Show me what happens when two users claim the last seat at the same time."

☝️ click to open app demo video

name = "Lost Update"
description = "Two transactions read and write the same row — only one update survives"
drop_tables = ["accounts"]

[[steps]]
sql = "CREATE TABLE accounts (id INT PRIMARY KEY, owner TEXT, balance INT)"
setup = true

[[steps]]
cmd = "begin"
trx = "alice"

[[steps]]
cmd = "begin"
trx = "bob"

[[steps]]
sql = "SELECT balance FROM accounts WHERE id = 1"
trx = "alice"

[[steps]]
sql = "SELECT balance FROM accounts WHERE id = 1"
trx = "bob"

[[steps]]
sql = "UPDATE accounts SET balance = balance - 100 WHERE id = 1"
trx = "alice"

[[steps]]
sql = "UPDATE accounts SET balance = balance - 100 WHERE id = 1"
trx = "bob"

[[steps]]
cmd = "commit"
trx = "alice"

[[steps]]
cmd = "commit"
trx = "bob"

Sequence building blocks:

• setup = true — initialization SQL, hidden by default (press s to show)
• cmd = "begin" / "commit" / "rollback" with a trx label — transaction lifecycle
• sql = "..." with a trx label — runs inside that transaction
• sql = "..." with no trx — auto-committed single-statement query

SQL sequences run with hidden responses, allowing you to test your understanding or quiz others on transaction behavior.

Press m or space to show responses.

Every request runs concurrently, with the UI showing when transactions wait for resource access.

Explore predefined sequences for common transaction scenarios.

The primary workflow pairs acid serve with an LLM agent (Claude, Gemini, or any agent that reads AGENTS.md) acting as a Socratic database coach.

Terminal 1                         Terminal 2
──────────────────────────────     ──────────────────────────────
make serve                         claude
  │                                  │
  │  acid serve starts on :7331      │  Agent reads AGENTS.md / CLAUDE.md
  │  Results hidden by default       │
  │                                  │  acid status          ← checks DB + server
  │  ┌─────────────────────┐         │  acid run -f sequences/lost_update.toml
  │  │ scenario plays out  │◄────────┤
  │  │ results are hidden  │         │  "What will the final balance be?"
  │  └─────────────────────┘         │
  │                                  │  [student answers]
  │  ┌─────────────────────┐         │
  │  │ results revealed    │◄────────┤  acid toggle
  │  └─────────────────────┘         │
  │                                  │  Debrief → next scenario

The agent writes new .toml files in sequences/, runs them with acid run, waits for your prediction, then calls acid toggle to reveal the results. acid init scaffolds the full environment including a pre-loaded coaching prompt in AGENTS.md/CLAUDE.md.

Agent commands:

acid status                              # verify DB connectivity and server health
acid run -f sequences/lost_update.toml  # stream scenario to the serve TUI
acid toggle                              # reveal results after student predicts

You can also run built-in sequences by name:

acid run "Lost Update"
acid run "Dirty Read"
acid run "Phantom Reads"

# macOS (Homebrew)
brew install --cask rusinikita/tap/acid

# Linux / macOS (shell script)
curl -fsSL https://raw.githubusercontent.com/rusinikita/acid/main/install.sh | sh

acid init my-learning
cd my-learning

This creates .env, AGENTS.md/CLAUDE.md (AI coaching prompt), learning_plan.md, a sequences/ folder with example TOML files, and a Makefile.

init command creates docker-compose.yml and Makefile for convenience.

make pg       # PostgreSQL in Docker (matches .env default)
# or
make mysql    # MySQL in Docker

Or edit .env to point at an existing cloud database (neon.com for PostgreSQL, planetscale.com for MySQL).

FIRST    make serve     # starts acid serve on :7331
SECOND   claude         # or: gemini, or any LLM agent

Say "Let's start" — the agent verifies the server is up, then kicks off the first scenario.

acid run -f sequences/lost_update.toml
acid toggle    # reveal results

• PostgreSQL
• MySQL

We welcome contributions! Here's how you can help:

1. Report Issues - Found a bug or have a feature request? Create an issue
2. Share Sequences - Create interesting transaction scenarios as TOML files
3. Improve Documentation - Help make the README clearer
4. Code Contributions - Submit pull requests for bug fixes or features

GPL-3.0 License - see LICENSE file for details.