Developer tools · MongoDB
mongoui: a local-first MongoDB GUI that runs AI on your explain() output, not your data
An AI-native Compass alternative that explains slow queries, measures the fix, and cleans data safely. The engine runs on your machine, so rows and credentials never leave it.
Inside the app
The real app, screen by screen
Real screenshots of the desktop app against a live database. The engine runs on your machine, so the rows never leave it.
Smart Scan
One click runs a full health pass across indexes, storage, duplicates, and schema.
Query optimizer
Your real explain() read back in plain English, with the index that fixes it measured on a sample.
Safe cleanup
Every change is a dry-run first: the exact blast radius and a before/after, with one-click undo.
Duplicate finder
Group by the fields that define a match, keep one of each, never a blind delete.
Schema doctor
Type drift, always-null, sparse, and unbounded-array fields, found before they bite.
Index & storage
Where storage actually goes, collection by collection, with reclaimable space called out.
The story behind
A MongoDB GUI is easy until the query is slow. Browsing documents is the simple part. The hard part is the aggregation that is slow for a reason the tool will not tell you. Compass shows what a query runs, not why. mongoui reads your real explain() output back in plain English: the COLLSCAN, the documents examined versus returned, and the specific index that fixes it. For an aggregation, where a pipeline cannot be paused server-side, it replays each progressively longer prefix, ends each one in a count so only a number comes back, and reports the real per-stage document deltas in the order you wrote them. It also flags when the planner reordered your stages, so explain() stops being a surprise.
The load-bearing decision was where the connection runs. Production and self-hosted databases sit on private networks a cloud app cannot reach, and holding someone's connection string server-side is the wrong posture for a tool you trust with prod. So the engine runs on your machine and connects directly. The only thing any AI path may send is a redacted context: schema, field names, index definitions, explain() output, and pipeline text. Raw values are off by default, and on means type-masked, never the real row. The optimizer holds the same honesty line: a proposed index is labelled estimated until it is measured on a throwaway sample, because MongoDB has no hypothetical-index feature and a guessed number is worse than none. And because Mongo has no undo, every cleanup snapshots the matched set before it writes, so a wrong update is one click to reverse.
Business value
- One click runs a full health pass over indexes, storage, duplicates, and schema, so problems surface before they bite.
- Production and self-hosted databases work, because the connection never leaves your machine.
- The AI answers from your real schema and explain() output, so it stops inventing syntax that looks right.
- A proposed index is measured on a sample before you trust it, not guessed.
- A bad-data cleanup carries a pre-write snapshot, so a wrong update is one click to undo.
Project scope
- A local engine with connection and topology detection, schema inference, and paged document browsing.
- A one-click Smart Scan health pass across indexes, storage, duplicates, and schema.
- An explain() parser and optimizer that proposes an index and measures it on a non-destructive sample.
- A prefix-replay pipeline debugger with real per-stage counts and planner-reorder detection.
- A safe-cleanup write path with a dry-run diff, a blast-radius count, and a snapshot rollback before any write.
- A redaction boundary that assembles the only payload any AI path is allowed to send.
Deliverables
- A signed and notarized desktop app for macOS and Windows, with Smart Scan, Browse, Optimize, Cleanup, Indexes, Duplicates, Space, and Schema doctor.
- A typed IPC bridge, so the UI never touches the driver or credentials.
- Claude Opus 4.8 narration that cites the real plan, the COLLSCAN, and the per-stage counts.
- A headless CI gate that fails a pull request when a pipeline regresses, on documents examined, not wall-clock.
Tech stack
Frequently asked
Is mongoui another admin panel?
No. It is a local-first desktop GUI focused on the hard part: reading your real explain() output, measuring the index that fixes a slow query, and cleaning bad data with a rollback. The connection runs on your machine, so production and self-hosted databases work.
Does my data get sent to the AI?
Not by default. The only thing any AI path may send is a redacted context: your schema, field names, index definitions, explain() output, and pipeline text. Raw sample values are off by default, and on means type-masked, never the real row. Bring your own key and the call goes straight from your machine.
How is it different from MongoDB Compass?
Compass shows what a query runs. mongoui works on why it is slow and how to fix it: an index measured on a sample before you trust it, per-stage document counts from a real pipeline replay, and AI grounded in your actual plan instead of generic syntax that looks right.
Can you build something like this for us?
Yes. The local-first architecture, the on-device redaction boundary, and the measure-before-you-trust pattern are things we can put on your own product or infrastructure.
Have a workflow that needs this?
Tell us the shape of the problem. Scoped estimate, usually within 3 to 5 business days. No card, no obligation.
Estimate this buildor email business@highcraft.ioMore work
AI Lab Analysis
AI-assisted intake reads raw lab PDFs and drafts plain-language summaries a provider reviews before the visit.
EMR & Chart Notes
An EMR and patient portal built in .NET and Azure, developed under real regulatory scrutiny on a normal sprint cadence.
Clinical AI Insights
Insights compiled from a patient's recent notes, latest lab, and intake, with the sources it used and the ones it could not.