SQL performance tuning for SQL Server · PostgreSQL · MySQL · MariaDB
Culprit reads the statistics your database already keeps and tells you, in one sentence, what's slow, what it costs and how to fix it.
Read-onlyNothing leaves your machineFree, MIT licensedmacOS · Windows · LinuxOther platforms ↓
order_items times out on every run, at about 15 s.order_items 1.5M times, once per row: the query forces a loop join, where one hash join over a single pass would do.OPTION (LOOP JOIN) and let the optimizer choose.order_items.Findings
Runaway loops in plans, plan regressions, parameter sniffing, missing and unused indexes, N+1 query storms, lock pile-ups, implicit conversions that block index seeks, spills to disk and stale statistics. Each finding comes with the evidence, a production-safe fix and how to verify it.
Plans
Insights in plain words, a diagram that shows where the cost goes, and every operator's details one click away. Compare any two plans, from SSMS, EXPLAIN or MySQL, and the headline says why one is slower.
What changed
Compare two periods and Culprit lists what got slower, what's new and what improved, next to the deploys and statistics updates around them: "Got 293× slower at 5:00 PM, when SQL Server switched from plan 163 to plan 198."
Live
A wall of lock waits becomes one sentence and a tree: which session holds everyone up, why, and who waits behind it. Deadlocks get the same treatment, with the pattern named and the fix spelled out.
Why Culprit
Every finding is ranked by its share of database time and comes with the evidence, a fix that's safe on production and a way to prove it worked.
A 14 MB native app that opens in under a second and shows the last report at once while a fresh scan runs in the background. Nothing to install on the server.
MIT licensed. Add your own checks or database engines without forking: the app, the CLI and the agent tools share one Rust core.
Built for coding agents
culprit mcp gives Claude Code, Codex and Cursor 25 tools: the dashboard, findings, execution plans and their insights, what changed, blocking, waits, index advice and verify. Every CLI command speaks JSON too.
culprit audit.
$ claude mcp add culprit -- culprit mcp
$ culprit agent init # set up this repo's agents# ~/.codex/config.toml
[mcp_servers.culprit]
command = "culprit"
args = ["mcp"]// ~/.cursor/mcp.json
{ "mcpServers": { "culprit": { "command": "culprit", "args": ["mcp"] } } }> the order report keeps timing out on staging. why?
● culprit.dashboard staging
● culprit.findings staging · severity ≥ medium
● culprit.plan_insights query 263
It times out on every run (about 15 s, 27% of database time): its plan repeats a read of order_items 1.5M times because the query forces a loop join. Removing OPTION (LOOP JOIN) lets SQL Server hash-join once.Safe on production
Scans use catalog and statistics views with short lock timeouts. Estimated plans never run anything.
If Culprit can prove a statement only reads, it runs it read-only. Anything else on production asks you first.
Every fix is shown and copied, never applied, and uses online or concurrent operations where the server has them.
No account, no cloud, no telemetry. Passwords live in your OS keychain; Entra ID uses your az login.
Supported databases
Query Store, the plan cache, DMVs, wait statistics and the deadlock history in Extended Events.
Versions 15 to 18: pg_stat_statements, the statistics views and EXPLAIN, generic plans included.
Versions 5.7 to 8.4: performance_schema, sys and EXPLAIN FORMAT=JSON.
Versions 10.5 to 11.8: performance_schema, the slow query log and ANALYZE FORMAT=JSON.
Download
Free and open source under the MIT license.
Apple silicon and Intel · macOS 11+
Signed and notarized by Apple
Windows 10 and 11 · x64
Not code-signed yet: SmartScreen may ask once
x86_64
AppImage or Debian package
One binary for scripts, CI and agents. Download it for macOS, Linux or Windows, then:
$ culprit conn add shop --url 'postgres://app@db.example.com:5432/shop'
$ culprit dashboard shop # the database at a glanceAll releases and checksums: github.com/cristipufu/culprit-sql-studio/releases
Read access to the server's statistics. SQL Server needs VIEW SERVER STATE (VIEW DATABASE STATE on Azure SQL Database). PostgreSQL needs pg_read_all_stats or pg_monitor. MySQL and MariaDB need PROCESS and SELECT on performance_schema. When a check can't run with your login, Culprit names the exact grant that unlocks it.
No. There's no agent or collector: Culprit connects like any client and reads the statistics the database already keeps (Query Store, pg_stat_statements, performance_schema). If one of them is switched off, it shows the one statement that turns it on.
No. Scans only read statistics and catalog views. The SQL editor runs reads read-only, and anything that could change data on a production connection asks you first. Suggested fixes are shown, never applied.
It's an investigation tool: open it when something is slow, or let your coding agent ask it. History comes from Query Store and from Culprit's own saved scans, so there's no server to run and nothing collecting around the clock.
Yes. Azure SQL Database and Amazon RDS for PostgreSQL are tested, Entra ID sign-in uses your az login, and Culprit can find your Azure and AWS databases with your existing CLI login.
Those show the numbers. Culprit reads them for you: it ranks what costs the most time, explains it in one sentence, gives the fix and tells you how to verify it, with the same views on all four engines.
Any agent that speaks the Model Context Protocol, including Claude Code, Codex and Cursor. Agents can also call the CLI directly, since every command has JSON output.
Yes. Culprit SQL Studio is open source under the MIT license: the app, the CLI and the MCP server.