Business Intelligence. Powered by AI.

Ask in English. Get SQL back.

Qrly compiles natural-language questions into QQL/SQL via your chosen LLM — Claude, Gemini, OpenAI, or local Ollama. Multi-turn conversations, schema-aware, never auto-runs destructive SQL.

Alert that enforces itself.

Configure threshold and goal-based alerts per question. Triggers fan out automatically — retrying, escalating, delivering to Slack, Teams or webhook — without human intervention.

Multi-tenancy from day one.

The entire data model is built around Tenant → Organization → Project. Every feature — auth, AI, mailboxes, webhooks — is scoped at the correct level with zero data leakage.

Subscriptions as a first-class citizen.

Not a plugin. Not a webhook hack. Native scheduled subscription pipeline with four delivery channels, cron schedules, signed embed tokens, and per-recipient format selection.

AI that stays in its lane.

Claude, Gemini, OpenAI, Ollama, LM Studio — all configurable per org. AI accelerates Ask (NL → SQL), schema descriptions, chart suggestions and anomaly detection, with hard rate limits, circuit breakers, and full usage logging.

AI woven into every workflow.

Qrly's AI layer is provider-agnostic — Anthropic Claude, Google Gemini, OpenAI, Azure OpenAI, Mistral, Ollama, LM Studio, Jan.ai, LocalAI and anything else speaking the OpenAI protocol. Configure per organisation. Encrypted keys. Per-user daily token budgets. Local execution for full data sovereignty.

• Ask — natural language → QQL, compiled, executed and narrated
• Schema descriptions — AI-written docs for tables & columns, schema-only by default
• Performance analyzer — a heuristic score with findings ranked HIGH to INFO; AI recommendations are an explicit step, so no tokens are spent by default
• Anomaly detection — |z| > 2.5 and week-over-week deltas above 25%, surfaced in the Insights feed
• Chart suggestions — picks the right viz for the result shape
• Filter suggestions — recommends interactive parameters and their lookup SQL
• Context drawer — opens the exact system prompt that was sent, broken down by source

QQL across any database.

One JSON query document compiles to PostgreSQL, MySQL, MariaDB, SQL Server, BigQuery, Snowflake, Redshift, SQLite, H2, DB2, Sybase and Pervasive — with Trino, Presto, DuckDB, ClickHouse, Vertica, Athena and Spanner folding into a compatible dialect. Roughly 35 operators across regex, JSON path, range, full-text and arrays. Visual builder, raw SQL editor, OLAP models — three origins, one engine.

• 40 connection types across 12 dialect compilers, including managed AWS, Azure and GCP variants
• 13 chart types: bar, line, area, pie, scatter, KPI, gauge, progress, funnel, pivot, maps — plus TREND on dashboards
• OLAP star/snowflake models with ROLLUP, CUBE, GROUPING SETS, pivot and prior-period comparison
• Live dashboards: WebSocket interval or PostgreSQL LISTEN/NOTIFY on the data source
• Interactive filters, calculated fields, forecasting and drill-through
• Question-as-view — reuse any saved question as a source or join target, five levels deep, with cycle detection

Any Identity. One Platform.

Run multiple auth providers simultaneously. Employees use Azure AD, contractors use a password, partners use Google — all on the same organisation, and the login page decides which buttons to show from the hostname it was reached on.

• Azure AD and Google (OAuth2/OIDC) — platform-wide clients and per-organisation registrations, composed at runtime
• LDAP and Microsoft Active Directory per organisation, with group sync and encrypted bind credentials
• SSO discovery by hostname — a custom domain resolving to one organisation narrows the page to that organisation
• API tokens qrly_live_…, SHA-256 hashed, scoped, rate-limited per token, revealed once
• JWT in an HttpOnly cookie with a denylist for logout and revocation, rotating keys and a validation cache
• Configurable session timeout, max sessions per user, login-failure tracking and per-IP lockout

Four personas. Nothing writes unasked.

An Analyst that answers, a Composer that builds dashboards, a Modeler that proposes OLAP models and an Investigator that runs a fixed six-step root-cause method with every write tool stripped out of its prompt. 29 tools — 16 read, 13 write — and every single write stages a proposal for a human to approve or reject.

• Propose then approve — write tools have no path to the real service call; approval required defaults to on
• Four stop conditions — step cap, daily token budget re-queried mid-run, wall-clock deadline, provider error
• Repetition circuit breaker — the same call three times returns a synthetic observation telling the model to change approach
• Server-side tool allow-list per session, on top of prompt-level filtering
• Full audit — every tool call with step, input, output, duration and status; opt-in per-step prompt auditing
• Verified findings — an independent validator re-checks every outlier the Investigator claims

Four cache tiers. Then replication.

Warm results in the JVM, persist them across restarts, materialize them as real tables in your own warehouse, or replicate the source tables outright over change data capture. All opt-in per connection or per question, all encrypted, and every cache failure degrades to a live query rather than an error.

• T1 JVM cache — in-process, byte-budgeted, with an optional columnar store of typed primitive arrays
• T2 persistent cache — the application database or a per-connection SQLite WAL file, laid out per tenant and organisation
• T3 materialized tables — a real table in a database you own, directly queryable by any other tool
• T4 table sync — full reload, watermark, snapshot-delta or CDC through embedded Debezium on Postgres, MySQL and SQL Server
• AES-256-GCM at rest in both cache tiers — the cache never stores plaintext rows
• Read routing transparently rewrites visual questions to the synced copy; native SQL is never rewritten

A real workbook. Bound to live SQL.

Roughly 500 worksheet functions with their own tokenizer, parser and evaluator — named ranges, dynamic-array spilling, cycle detection and the full Excel error set. Then bind a block of cells to a saved question, stored QQL, ad-hoc SQL or an AI prompt, and refresh it without destroying the work built around it.

• Refresh-safe blocks — source columns are addressed by name, so inserted columns and rows are preserved and never overwritten
• Key columns return each record to its previous row, so hand-written cells stay beside their record across reorders
• AI bindings resolve once to SQL, then refresh as plain SQL — a scheduled refresh never spends tokens
• 17 conditional-format presets plus colour scales, data bars and icon sets with custom thresholds
• Point mode, fill handle, paste-special including transpose, frozen panes, merged cells, find and replace
• Server-side export to XLSX with real formulas, CSV, JSON and XML — the same bytes a subscription emails