Introduction
DeepSQL is an open source database agent for Postgres and MySQL. It watches your workload and suggests indexes. It also rewrites slow queries and answers business questions in plain language. You install it inside your own infrastructure with a one-line curl script, and per its pricing page the full product costs $0 with no feature gates, seat limits or usage metering.
That is an appealing pitch for anyone who has been both the person tuning EXPLAIN output at midnight and the person fielding "can you pull the signup numbers" requests from the sales team. This review covers what DeepSQL's own pages and repository say it does, what it costs, and where the gaps are. scored.tools has not run it against a live database. Every capability below comes from the vendor's site or the public GitHub repo, and we attribute it that way.
Our overall score is 6.2/10. The pricing and install story are strong. The weak spots are the thin independent evidence on output quality and the early version number.
Key Features
Workload monitoring and index advice
The homepage and pricing page describe 24/7 workload monitoring, index and materialized view recommendations, unused index and bloat detection, and explain plan analysis. The idea is that DeepSQL learns your actual query patterns and proposes changes from them. The homepage's sample terminal output shows recommendations waiting for review rather than being applied automatically. For a tool that touches production schemas, a human review step is the right default.
Slow query detection and rewrites
DeepSQL says it detects slow queries and proposes rewrites. There are no published before-and-after benchmarks from independent users, so how good those rewrites are is still an open question. The vendor shows sample outputs. A bad index suggestion costs you write throughput and disk, so treat each recommendation as a draft to check.
Natural-language BI and dashboards
The second half of the product is a BI layer. Non-technical users ask questions in plain language, and DeepSQL generates the SQL and can build dashboards plus scheduled, shareable reports. Per the homepage, the agent works through a visible sequence. It resolves the schema, drafts the SQL, validates it, then runs it on a read replica. You can inspect each step, which matters more than any single answer, because you can see when it joined the wrong table.
Read replica execution and access controls
Queries run against a read replica using read-only credentials (the homepage example uses a deepsql_ro role) over VPC-peered connections with SSL required. The site also lists policy-enforced row and column access, PII masking in generated SQL, and audit logs. Running on a replica keeps analytical load off your primary. You do need a replica to get that benefit, which smaller setups may not have.
The Brain
DeepSQL calls its context layer "the Brain". It indexes your schema and workload, stores metric definitions and company rules, and can take in application code context. Each connection needs to finish a Brain initialization step (via deepsql connections init) before the agent gives full answers. Expect some lag between connecting a database and getting useful output.
Schema change tracking
A "Guardian" feature flags risky or recent schema changes. The homepage presents it as a prompt asking what changed recently. If you lack migration review discipline, this could catch drift. Teams with strict migration tooling will get less from it.
Interfaces
You can reach DeepSQL from a web UI, a CLI or Slack. The install also sets up an MCP server, so Claude, Codex and Cursor can query it as a tool. The pricing page lists no public REST API or SDK, so deeper custom integrations are limited to MCP for now.
Pricing Breakdown
There are two tiers, and only one of them is software.
| Plan | Price | What you get |
|---|---|---|
| Open Source | $0 | Full product: agent, Brain, monitoring, query rewrites, index advice, BI dashboards, access controls. Self-hosted in your VPC. Community support on Discord and Slack. |
| Enterprise Edition | No published price ("It depends on scope") | Same software plus consulting: guided VPC install, Brain tuning, workload audit with a fix plan, custom dashboards, onboarding, private support, architecture reviews. |
The real cost of the open source tier is the compute you run it on plus your LLM bill. DeepSQL requires you to bring your own LLM keys. An agent that indexes schemas, monitors workloads around the clock and answers ad-hoc questions will make a lot of model calls, and the site gives no estimate of that spend. Budget for it and set provider-side limits before you point it at a busy database.
There is no hosted SaaS option. The pricing page says DeepSQL always runs inside your own infrastructure. Security teams will like that. Teams with no ops capacity will not.
Scores
| Criterion | Score | Basis |
|---|---|---|
| Output quality | 4/10 | Only vendor sample outputs. The public repo is small (40 stars, 2 forks, 27 open issues), too little to judge output quality. |
| Ease of use | 8/10 | One-line curl install, about 15 minutes to self-host per the site, free to start. No independent usability reports. |
| Pricing value | 8/10 | $0 with no feature gates or seat limits. You pay for your own LLM usage. |
| API and integration | 6/10 | MCP server, web UI, CLI, Slack and multiple LLM providers. No documented public API or SDK. |
| Problem fit | 5/10 | Vendor lists the right DBA features. No user reports of results yet. |
| Overall | 6.2/10 |
Pros and Cons
Pros
- Fully open source and self-hosted, so query data and schemas stay in your environment.
- One agent covers DBA tuning and BI questions, which saves you from running two separate tools.
- Queries run on a read replica with read-only credentials, keeping load and risk off the primary.
- Visible step-by-step SQL generation lets you audit what the agent did.
- Works from the web, the terminal, Slack and MCP-aware coding tools.
- No seat limits, so adding analysts costs nothing beyond LLM usage.
Cons
- Version v0.26.0 on the homepage. This is pre-1.0 software.
- Independent evidence is thin. The GitHub repo had 40 stars and 27 open issues when we checked the data.
- Brain initialization per connection adds setup time before answers are complete.
- Postgres and MySQL only. No SQL Server, Oracle, Snowflake or other engines listed.
- LLM costs are on you and unestimated.
- No hosted option, and Enterprise pricing is unpublished.
Who Is It For
DeepSQL fits a small to mid-size engineering team running Postgres or MySQL (including RDS, Aurora or Cloud SQL, per the homepage) without a dedicated DBA. Those teams are often the ones whose analysts and product managers queue up SQL requests for backend engineers. If you already have a read replica and someone comfortable running a Linux service in your VPC, the cost of trying it is close to zero.
It is a poor fit if you have no ops capacity and want a managed service, or if your data lives in a warehouse rather than an OLTP database. Teams on Databricks already have their own analytics tooling. DeepSQL also does not replace a managed database platform such as Supabase or PlanetScale. It sits on top of the database you already run, so compare it with the observability and query advice those platforms ship rather than treating it as an alternative to them.
Regulated teams should look at the read-only roles, PII masking and audit logs with interest, then confirm them in the code. The project is open source, so you can.
Verdict
DeepSQL puts index advice, slow query tuning and natural-language BI into one open source agent, and it makes sensible design choices for something that writes SQL. Execution happens on a read replica, credentials are read-only, recommendations wait for review and every step is visible. The $0 price with no gates is real, apart from your LLM bill.
The gap is evidence. Every quality claim comes from the vendor, the repo is small, and v0.26.0 tells you the project is early. Our recommendation is to install DeepSQL against a staging database or a non-critical replica, let the Brain finish initializing, and compare its index suggestions with what your own EXPLAIN ANALYZE work shows before you trust it near production. If the suggestions hold up, it is a cheap way to give a team without a DBA some DBA-level help. If they don't, you have lost about an afternoon. It scores 6.2/10 today, with room to climb once users publish independent results.