SkillAgentSearch skills...

LiteDB

Local-first SQL client for SQLite and PostgreSQL where AI agents work through guardrails: every write is previewed, approved in the app, logged and undoable. Built-in MCP server, Text-to-SQL with local models, schema visualizer. Tauri + Rust + React.

Install / Use

claude mcp add createdbyadham -- npx -y github:createdbyadham/LiteDB

If the server publishes to npm under a different name, use that package instead β€” check the repo README.

About this skill
πŸ”Œ

MCP Server

Model Context Protocol server

Quality Score

84/100

Supported Platforms

Claude Code
Claude Desktop
Cursor

Our assessment of LiteDB

LiteDB scores 84/100 on our quality scale, 601st of 937 AI & Machine Learning skills we index.

Its MCP Server is 33 KB long, well organised into 30 sections with 14 code examples: a thorough specification that gives an agent plenty to work with.

It has 10 GitHub stars, so there is little community track record yet; judge it on its content.

Substance
30/30
Structure
20/20
Description
15/15
Adoption
4/20
Freshness
15/15

Maintenance, license and trust

  • The repository was last updated today, so LiteDB is actively maintained.
  • It is released under the MIT license, a permissive license that allows use, modification and commercial use with attribution.
  • Its trust signals score 97/100, with no cautions. These come from repository metadata, not a code audit β€” read the skill file before letting an agent act on it.

Safety scan

No issues found

Our scan of the whole file found no instruction hijacking, hidden characters, credential access, data exfiltration or destructive commands. An AI review of the same text found nothing harmful.

AI review by kimi-k2.7-code on 2026-11-18. Automated pattern scan on 2026-11-18. It catches known dangerous patterns, not every risk β€” read a skill before letting an agent act on it.

LiteDB compared with similar skills

All 4 of these similar skills score higher than LiteDB; compare them before choosing.

SkillScoreStarsUpdatedFormat
LiteDB (this skill)by createdbyadham8410todayMCP Server
claude-memby thedotmack100118.2ktodayCLAUDE.md
Agent-Reachby Panniantong100113.0k10d agoCLAUDE.md
Understand-Anythingby Egonex-AI10093.2ktodayCLAUDE.md
hyperframesby heygen-com10086.4ktodayCLAUDE.md

Frequently asked questions

How do I install LiteDB?
Run claude mcp add createdbyadham -- npx -y github:createdbyadham/LiteDB. The install tabs above show the steps for each supported agent.
Which AI agents does LiteDB work with?
It is written for Claude Code, Claude Desktop and Cursor, as a MCP Server file. Other agents that read the same format can often use it too.
Is LiteDB safe to use?
Our scan of the whole file found no instruction hijacking, hidden characters, credential access, data exfiltration or destructive commands. An AI review of the same text found nothing harmful. It is MIT-licensed and scores 97/100 on trust signals. Skills are instructions an agent will follow, so read the file before installing it and do not approve commands you do not understand.
Is LiteDB still maintained?
The repository was last updated today, so LiteDB is actively maintained.

LiteDB

A desktop database client for SQLite and PostgreSQL that lets AI agents in without letting them loose. When Claude Code, Cursor or any MCP agent wants to change your data, LiteDB shows you the statement, the exact row count and the rows it changes, and nothing runs until you click Approve. Change your mind later, and Undo puts the rows back.

It is also an everyday client: a real SQL editor with Text to SQL, inline editing, saved connections, SSH tunnels and production tags.

<div align="center">

Download for macOS Download for Windows Download for Linux

</div>

Claude Code asks to change 13 customers; LiteDB shows the rows and waits for Approve, the audit log names the agent, and Undo puts the rows back

Claude Code asks to move 13 customers from FR to DE. LiteDB shows exactly which rows and waits; after Approve, the audit log names the agent, and one click puts the rows back. (Video version.)

If this would make you trust an agent with your database, a ⭐ helps others find it.

Install

Download the installer for your platform from the latest release:

Homebrew, winget, Scoop, AUR and Flathub packages are on the way; until they are listed, use the downloads above.

The macOS app is signed and notarized by Apple. The Windows installers are not code-signed yet, so SmartScreen may warn: choose More info, then Run anyway.

Features

  • SQLite, opened in place: files are read and written natively (rusqlite), so a multi-gigabyte database opens as fast as a small one and shares the file with your app or an agent through SQLite's own locking. Writes from another program show up in the grid on their own.

  • PostgreSQL the way it is used: every schema the role can see (Supabase's auth and storage included), paste a postgres:// URL to fill the form, SSL from off to verify-full with your CA, and SSH tunnels through a bastion with a key, a password or your SSH agent.

    On Supabase, paste the connection string from the project's Connect panel and type the password. The panel shows [YOUR-PASSWORD] in its place. Prefer the Session pooler string (port 5432) for a desktop app. The direct connection (db.<ref>.supabase.co) is IPv6-only unless the project has the IPv4 add-on. The Transaction pooler (port 6543) works, but SET and temporary tables do not last between runs, and Cancel can only stop waiting, not stop the server.

  • Saved connections: as many as you like, each with a name and a colour. Passwords and SSH passphrases go to the OS keychain, only if you ask.

  • Production tag: a connection tagged production opens read-only, for you and for any agent following the app, and turns the title bar red while it is on screen. Leaving read-only asks first.

  • Several databases open at once, as tabs in the title bar. Switching does not reconnect.

  • SQL editor: syntax highlighting, autocomplete that offers the columns of the tables in your query (no table. needed), Ctrl/Cmd+Enter runs the statement under the cursor or the selection, F5 runs the script, query tabs that survive a restart, history and saved scripts, a row limit, and Cancel for a long query. Errors underline the statement (and, in Postgres, the exact spot).

  • AI in the editor: Ask AI (Ctrl/Cmd+K) opens Text to SQL from the editor, Explain describes a query with its plan, and a failed query offers Explain and Fix with AI, shown side by side before anything changes.

  • Inline suggestions (off by default): the model's continuation as grey text after the cursor, Tab to accept, Esc to dismiss. Fill-in-the-middle on Ollama, chat elsewhere. Never on a production connection, and accepted text counts as AI-written for the safety gate.

  • Table editor: double-click a cell to edit it, insert and duplicate rows, filter by any column, follow a foreign key to the row it points at (or see what points here), view and edit JSON, and copy rows as TSV, CSV, JSON, INSERT statements or Markdown. Results scroll smoothly however many rows come back.

  • CSV, Excel and JSON files: open a .csv, .tsv, .xlsx, .xls, .ods, .json or .jsonl file like a database, a table per file, sheet or JSON array, with column types read from the data. Browse it, query it with SQL or Text to SQL, and export it. It is a read-only copy; the file is never written.

  • CSV import: into a new table with inferred column types, or into an existing one with columns matched by name, in one transaction.

  • Command palette: Ctrl/Cmd+P jumps to any table, connection or action.

  • Light and dark themes, or follow the OS.

  • Schema Visualization: Visualize your database structure, relationships, and foreign keys in an interactive diagram.

    • Auto-Layout: Automatically arranges tables to minimize crossing lines.
    • Export: Save your schema diagram as PNG or SVG for documentation.
  • AI Agent (Text-to-SQL): Turn natural language into SQL queries.

    • Settings β†’ AI: OpenAI, Anthropic (Claude), Google Gemini, OpenAI-compatible (LM Studio, vLLM, GitHub Models, Azure, …), or Ollama. Default is Ollama.
    • Same prompt the eval harness measures: schema + sample values (toggle), retrieved few-shot (K=3), one compile-repair pass.
  • MCP: Settings β†’ Agents, or the status-bar MCP chip. Claude Desktop / Claude Code / Cursor / OpenCode / VS Code follow the open connection, Connect an agent.

  • Write Safety: Generated SQL is classified before it runs.

    • Read-only / guarded / unrestricted / YOLO modes, remembered per connection.
    • Unqualified DELETE/UPDATE, DROP and TRUNCATE intercepted.
    • Approval shows the exact affected-row count, not just a warning.
    • Generated SQL is compiled before you see it, so the model fixes its own unknown columns and syntax errors, writes included.
    • Append-only audit log of every statement that ran, including edits, inserts, deletes and imports made in the grid, and who authorised it.
    • An agent's writes wait for your Approve in the LiteDB window, with the rows they change, and can be undone from the audit log.
  • Export: tables or query results to CSV, Excel, JSON or INSERT statements.

  • Vector Search & Semantic Search: Perform semantic similarity searches on your data using pgvector and local embedding models.

Measured, not asserted

LiteDB's Text-to-SQL agent ships with a public benchmark in evals/ , 109 cases scored by execution accuracy, with a held-out split that was written after the prompt work and never tuned against.

qwen2.5-coder:7b running locally on a 6 GB laptop GPU, SQLite:

| | Overall | Held-out | | --- | ---: | ---: | | Schema context + execution-guided repair | 77.3% | 67.6% | | + retrieved few-shot exemplars | 85.6% | 76.5% |

On a second, unseen schema, different naming conventions, integer cents, nullable dates, the same agent scores 91.7%, so the gains are not an artifact of one database.

npm run eval:selftest   # verify the harness itself (no API calls)
npm run eval -- --provider ollama

Six experiments are documented in evals/README.md, including the two that were rejected and the case-design defects a stronger model exposed. Every number is reproducible with one command.

That score is also the argument for the write-safety layer: 85.6% is good for a 7B model on read-only queries, and read-only queries are the easy case. Generated SQL is classified before it runs, and never executes unattended, even on a connection you have set to unrestricted.

Compared honestly

LiteDB is a SQLite/Postgres desktop client with an agent and a write-safety layer. It is not a replacement for a general DBA workbench.

| | LiteDB | TablePlus | Beekeeper Studio | DBeaver | | --- | :---: | :---: | :---: | :---: | | SQLite + PostgreSQL | βœ“ | βœ“ | βœ“ | βœ“ | | MySQL, SQL Server, and the rest |, | βœ“ | βœ“ | βœ“ | | SSH tunnel / jump host | βœ“ | βœ“ | βœ“ | βœ“ | | ERD | βœ“ | paid | plugins | βœ“ | | Text-to-SQL (local or cloud) | βœ“ |, |, | Pro AI | | Write-safety (classify, row-count approval, audit) | βœ“ |, |, |, | | MCP server through those guardrails | βœ“ |, |, |, | | An agent's write waits for Approve in the app, with the rows it changes | βœ“ |, |, |, | | Undo for an agent's write | βœ“ |, |, |, | | pgvector semantic search in the app | βœ“ |, |, |, | | JDBC / every obscure driver |, |, |, | βœ“ | | Native polish / years of query UX |, | βœ“ |, |, |

The loss that matters in practice: two dialects. If you need MySQL or SQL Server, use those tools. Use this one when a model is allowed near the database, when production should be hard to change by accident, or when pgvector is the point.

Schema Visualization

LiteDB now includes a powerful Entity Relationship Diagram (ERD) generator:

  1. Interactive Graph: Drag and drop tables, zoom in/out, and explore relationships.
  2. Visual Foreign Keys: Lines connect Foreign Keys (Source) to Primary Keys (Target) automatically.
  3. Key Indicators: Visual icons for Primary Keys (πŸ”‘), Foreign Keys (πŸ”—), and Unique constraints (#).
  4. Export Ready: One-click export to high-quality images for your technical documentation.

Vector Search & Semantic Search

LiteDB integrates advanced vector search capabilities powered by pgvector and local embedding models:

  1. Semantic Search: Find similar rows based on vector embeddings.
    • Search by Row ID: Find rows that are semantically similar to a specific record.
    • Search by Text: Enter natural language queries to find relevant records using local embedding models.
  2. Local Embedding Models: Run embedding models locally in your browser/app using Transformers.js.
    • Supported models: all-MiniLM-L6-v2, bge-base-en-v1.5, bge-large-en-v1.5.
    • Privacy-first: No data is sent to external APIs for embedding generation.
  3. Distance Metrics: Support for multiple distance metrics to suit your data:
    • Cosine Distance (<=>): Best for normalized vectors.
    • L2 Distance (<->): Euclidean distance.
    • Inner Product (<#>): Dot product (negative).
  4. Visual Feedback: Color-coded similarity bars to quickly identify the most relevant results.

AI Architecture (Text-to-SQL)

Unlike standard API wrappers, LiteDB implements a Context-Aware RAG Pipeline to ensure high-accuracy SQL generation. Settings β†’ AI is the live matrix: OpenAI, Anthropic (Claude), Google Gemini, OpenAI compatible, Ollama (local). GitHub Models / Az

Truncated for display β€” read the full file on GitHub.

Related Skills

View on GitHub
GitHub Stars10
CategoryAI
Updated18h ago
Forks1

Languages

TypeScript

Trust signals

97/100

From repository metadata: license, adoption, age and documentation. Not a code audit β€” see the Safety scan above for what the skill file itself contains.

1 info
LiteDB β€” MCP Server: Install & Safety Check | SkillAgent