the full tour

Everything Squinty can do for you

Seven chapters, in the order you will meet them: ask, build, connect, feed, watch, trust, run. Every claim below is a shipped feature, not a roadmap slide.

ask

A conversation, not a query builder

why did on-time delivery drop last month?

I checked June against your usual pattern. One distribution center drove 41 late deliveries on the Gulf Coast lane. Want that as a chart, or should I set an alarm on it?

Studio · chat with Squinty

Plain-language answers: Counts, trends, top-N, distributions, sample rows. Explore first, build charts only when you want them.
Voice input: Dictate your question from the composer, with honest listening and thinking states.
Explain this dashboard: Ask about any page or widget on screen. Click a widget and ask about it without even naming it.
Clarifies instead of guessing: An ambiguous request gets a clarifying question back, not a confidently wrong chart.
Web context when useful: Squinty can pull public information like weather or prices and combine it with your internal numbers.
Chats and Projects: Every conversation is saved and searchable, and can be organized into projects. History is private to you.
Memory with consent: Squinty proposes what to remember on approval cards. Memories are encrypted per user, pinnable, exportable, and truly forgettable.
Daily briefing, opt-in: A morning summary built from your approved memory and pending suggestions. Never on by default.

build

From one sentence to a multi-page application

create a wholesale ops app with sales, logistics and quality pages

Built it: four pages, a menu, twelve widgets. Drag anything anywhere, your layout is saved just for you. Say the word and I will regenerate any single widget.

A generated application

A generated multi-page application with KPI tiles, charts, a table, and a heatmap

Real applications: Multi-page apps with menus and menu groups, not just a single dashboard canvas.
Nine widget types: KPI, table, line, bar, pie, histogram, multi-axis, heatmap, and text blocks, chosen automatically.
Revise by asking: Regenerate one widget, remove a series, add a page. Only what you asked about changes.
Your layout is yours: Drag-and-resize grid with per-user layouts. Colleagues viewing the same app keep their own arrangement.
Interactive charts: Zoom, toggleable legends, and one-click CSV export of exactly the data on screen.
Honest widgets: If a number cannot be grounded in your schema, the widget says unavailable with the reason. Never a plausible wrong answer.
Own it, share it: One owner per app, read-only sharing, request-access flow, and an inbox with approve, deny, and revoke.
Instant start: A default app is generated for first-time users before they write a single prompt, plus a guided tour.

connect

Point it at the databases you already have

connect our orders warehouse and keep it fresh

Connected and introspected: 214 tables, relationships mapped. I re-sync every schema daily at 05:00 UTC and warn you visibly if something changes or breaks.

Data Sources

PostgreSQL and Oracle, end to end: Fully supported for natural-language querying, schema introspection included. Connections also cover MySQL, MariaDB, SQL Server, and Snowflake.
Automatic schema understanding: Tables, columns, and join paths are introspected for you. No semantic layer to hand-build or babysit.
Daily sync with visible status: Every source is re-introspected daily, with last-synced stamps and plain-English failure reasons. A failed refresh never destroys the last good schema.
Read-only by construction: The query engine can only ever SELECT. No chat message can modify or delete anything in your database.
Safe deletion: Removing a source that applications depend on is blocked with a list of what depends on it.
Workspace Knowledge built in: Every user gets a private managed source holding structured data derived from their own files.

feed

Documents in, cited answers and whole apps out

here is the supplier contract, check our invoices against it

Read all 42 pages. Your March invoices use the old rates from appendix B, about $8,400 over. Answer grounded in the exact file versions, citations included.

Knowledge · Files

Knowledge library with files in Ready, Reading and Partially ready states

Drop files into the chat: PDF, Word, Excel, CSV, JSON, Markdown, text, and images. Attach with a paperclip or drag and drop, reading starts immediately.
Grounded, cited answers: Content is chunked and indexed semantically. Replies say which attached files were used as context.
Apps from documents alone: Generate a complete application from selected files with no SQL data source at all, with the exact file versions recorded in the app.
Honest reading status: Uploading, Reading, Ready, Partially ready, Failed, Encrypted, Corrupt. Nothing fake, nothing stuck at 99 percent.
Images too: With a vision provider configured, images are transcribed and described, then indexed and citable like any document. You can also ask ad-hoc questions about a screenshot.
Versioned and pinned: Files are versioned and immutable. An answer stays pinned to the version it actually used.

watch

It keeps watching after you close the tab

alarm me when late deliveries look abnormal, and send ops a Monday summary

Two things armed: a forecast-based anomaly alarm that knows your weekly rhythm, and a Monday 07:00 HTML report to the ops list. Both visible in Monitoring.

Monitoring · Alarms and Scheduled Reports

Monitoring panel with alarm policies and scheduled reports

Alarms in plain English: Describe the condition, Squinty writes the policy. Managed in the app's Monitoring panel.
Four detection modes: Fixed thresholds, adaptive percentile bands, z-score outliers, and forecast-based anomaly detection that learns trend and seasonality.
Email delivery that retries: Notifications track their own state with retry and escalation. Duplicate sends are impossible by design.
Scheduled reports: Text, CSV, or HTML, scoped to a whole app, one page, or a single widget, on any cadence.
Scheduled data ingest: CSV and Parquet datasets loaded on a schedule with mapping, de-duplication, and a dead-letter path.

trust

Built for the security review, not around it

our auditors will ask who saw what and when

Good. The audit log has the actor, the action, a before-and-after diff, IP, and correlation ID for every change. And no, admins cannot read anyone's private memories.

SSO and MFA: Google single sign-on via OIDC plus native accounts, with authenticator-app two-factor enforced before anything opens.
Isolation in the database itself: Row-level security enforces workspace boundaries below the application layer. Cross-workspace probing returns not-found, indistinguishable from nonexistence.
Owner-private by default: Apps, sources, chats, files, and memories belong to their creator. Access moves only through explicit sharing with approve, deny, and revoke.
Sub-workspaces: Delegated administration over a governed hierarchy. Siblings can never see each other.
Audit trail: Who did what and when, with full diffs and secrets masked out of every entry.
Server-enforced sessions: Sliding idle timeout and a hard absolute limit, enforced server-side, with honest expiry messages.
Governed plugins: Extensions are validated data packages rendered by the platform, not arbitrary code. Owners approve the catalog, admins toggle per workspace, revocation is instant everywhere.

run

Runs on our models, your models, or no cloud at all

legal says the schema cannot leave the building

Then it will not. Point my analyst lane at your own GPU box, Ollama or any OpenAI-compatible endpoint, with failover. The planner never sees SQL either way.

LLM Config

Bring your own model, per user: Each user configures provider, model, and endpoint for both AI lanes. Keys are envelope-encrypted per user and record.
Cloud or local: OpenAI, Anthropic, xAI, any OpenAI-compatible endpoint, or local runtimes like Ollama, LM Studio, and LocalAI.
Fully local operation: Air-gapped deployments run against self-hosted models end to end, for privacy, residency, and cost control.
Multiple local servers: Run several servers for the same model. If one goes down, the others take over.
No silent defaults: An unconfigured lane says so plainly instead of quietly billing a fallback.
Deploy anywhere: Docker Compose on one machine or Kubernetes with Helm on a cluster, same codebase, with distributed tracing built in.