The 'minimal' SQL client that gates your AI agent's writes
A fast desktop SQL client for Postgres, MySQL, SQL Server, and SQLite that doubles as an MCP server — AI agents can read your data freely but need your click to write.

What it does
data-peek is an Electron-based desktop SQL client for PostgreSQL, MySQL, SQL Server, and SQLite. Beyond the query editor (Monaco, autocomplete that understands table aliases, multi-tab), it covers schema browsing, inline data editing, a table designer, ER diagrams, CSV import/export, and a health dashboard that can kill runaway queries. An AI assistant turns plain-English questions into schema-aware SQL using your own key from OpenAI, Anthropic, Google, Groq, or a local Ollama model. There’s even a terminal companion (@data-peek/cli) that runs eight Postgres schema checks — missing primary keys, unindexed foreign keys, bloat — usable as a CI gate.
The interesting bit
It also runs an MCP server: your saved connections become tools an AI agent can call over local HTTP, secured with a bearer token and off by default. Reads are fenced — 500-row cap, rollback-wrapped, and Postgres connections run READ ONLY at the database level — while every write via execute_statement pops an Approve/Reject dialog in the app and auto-rejects after 60 seconds. Human-in-the-loop for agent writes, built into the client rather than bolted on.
Key highlights
- Watch Mode pins a SELECT and re-runs it on a cadence from 500ms to 5 minutes, flashing changed cells amber — and flatly refuses to poll anything mutating.
- Time Machine snapshots every successful SELECT locally (50 runs per query, 512 MB cap), so you can scrub back through past results and diff any two runs cell by cell.
- Cross-tab references: name a query tab, then use
@namein another tab; the result gets inlined as a CTE at run time. - Benchmark mode reports p50/p90/p99 across repeated runs, alongside an EXPLAIN viewer and automatic detection of missing indexes and N+1 patterns.
- Credentials live in the OS keychain, there’s no telemetry, and the optional audit log is hash-chained and tamper-evident.
Caveats
- “Minimal” is doing heavy lifting: the feature list spans fake-data generation, trigger management, and a marketing site, and the sub-2-second startup claim is self-reported with no benchmarks in the README.
- The bring-your-own-agent mode (Claude Code, Codex, Antigravity) fully grounds against your live database only with Claude Code; the others wait on upstream headless tool approval.
- The audit log stores raw SQL text, which can contain data values — it’s off by default for a reason.
Verdict Worth a look if you live in a SQL GUI all day and want something snappier than the incumbents, or if you want to hand an AI agent read access to your database with a human gate on writes. Skip it if you need anything beyond the four supported databases, or if Electron is a dealbreaker for you.
Frequently asked
- What is Rohithgilla12/data-peek?
- A fast desktop SQL client for Postgres, MySQL, SQL Server, and SQLite that doubles as an MCP server — AI agents can read your data freely but need your click to write.
- Is data-peek open source?
- Yes — Rohithgilla12/data-peek is an open-source project tracked on heatdrop.
- What language is data-peek written in?
- Rohithgilla12/data-peek is primarily written in TypeScript.
- How popular is data-peek?
- Rohithgilla12/data-peek has 1.7k stars on GitHub.
- Where can I find data-peek?
- Rohithgilla12/data-peek is on GitHub at https://github.com/Rohithgilla12/data-peek.