What Curltale AI does
What you can do with Curltale AI, and why it exists.
| You want to... | Where | In short |
|---|---|---|
| Describe your data once | Ontology | Entities, fields (columns and calculated values), relationships, named filters, query examples, custom properties. Draft, review, publish, with an audit trail and optional four-eyes publishing. |
| Bring your own spreadsheet or database | Ontology > Changes & versions, Settings > Connectors | Fill in an Excel template, or point a connector (PostgreSQL, MySQL, MariaDB, ClickHouse, SQLite, DuckDB) at a database and generate a first draft from its structure. |
| Let agents ask questions of your data safely | MCP, Alfred | Plans for one correct SQL statement, run read-only under per-token or per-user entitlements; restricted columns never leave, personal ones are hashed. |
| Turn documents into reliable, cited knowledge | Knowledge | Upload files, capture web pages, or interview an expert. A model writes a verified source wiki for each document, a second model checks it claim by claim, and topics are written from the full wikis with quote-backed claims and typed relationships, all chunked, embedded and searchable. |
| Keep knowledge separate where it must be | Knowledge > Collections | Each collection is isolated with per-user and per-token grants and a public or internal usage tag that tells agents how to treat what they read. |
| Review before anything is published | Knowledge > Review | Contradictions, probable duplicates, pages the checker objected to, suggested model links, unreadable documents and vocabulary proposals wait in the collection's queue. |
| Connect documents to the model | Topic pages, entity pages, graph | Topics are linked to entities ("this policy governs Invoice"), shown on both sides, stored inside the collection and never in the model. |
| Give any agent both | MCP | Sixteen read-only tools: six for the model and the data, ten for documents (led by library_context, which returns whole topic pages with their graph; the document tools only appear for a token that can read a collection). The server explains itself, no prompt engineering. |
| Ask in plain language | Alfred | The in-app assistant uses the same tools and rules, in a side panel or full screen. |
| Try it without real data | Data > Browse | A synthetic database generated from your model, optionally with realistic values from the AI model. |
| Run it your way | Optional services | Separate worker, browser service for JavaScript pages, your own S3 bucket, any OpenAI-compatible AI endpoint, local embeddings. |
| See everything that happened | Admin > Activity | Every assistant conversation and agent session, step by step: model calls with tokens and cost, tool calls with timing and the response returned. |
Why it exists#
Language models write SQL well, but they do not know your business: which table is the customer, which of three status columns means "open", that two tables must never be joined flat because sums double, that a column holds personal data, or that the refund policy changed last spring. Without that knowledge, an agent guesses, and its answers look right and are wrong.
Curltale AI is that knowledge, written down once, kept current by the people who own it, and enforced for every consumer:
- One source of truth. Definitions live in one published model and in governed collections, not in prompts, dashboards and people's heads.
- Correct by construction. Agents receive the tables, join keys and governed conditions for a question as one plan, and the database does the joining and aggregating.
- Evidence for every fact. Document knowledge is verified before it is used: each document becomes a checked source wiki, topics are written from those wikis, and every fact is backed by a claim with its exact quote, the source and the wiki version it came from. Anything doubtful waits for a person.
- Safe by default. What a caller may see is decided by the server, column by column and collection by collection, for every query and every search, whoever asks.
- Any agent. Anything that speaks MCP (Claude, ChatGPT-based tools, custom agents, IDE assistants) can use it with a token.