One program you run yourself. It learns your business from your own files and answers in plain English — and underneath, it's an immutable, tamper-evident system of record. No cloud, no per-token bill, no license. Here's what that means for founders, the C-suite, and developers.
Drop in everything your business runs on and just ask. Centauri is the AI employee that already knows your stuff, costs nothing per question, and keeps every secret in the building.
Run centauri desktop — it detects your hardware, installs the AI runtime, and pulls the right models by itself. No cloud account, no configuration, no consultant.
"Which invoices from Acme are overdue?" "What did we promise this client?" It answers and shows you exactly where it found it.
PDFs, invoices, contracts, scanned images — a local vision model pulls out the facts so they're searchable, no manual data entry.
New data embeds itself the moment you add it, so search and answers are ready immediately — no setup, no separate tool.
Thumbs-up or down on an answer's sources and it re-ranks next time. It learns your business — without ever changing the underlying model.
The AI runs on your own computer. No subscription, no per-question charge, no vendor watching your data. (An optional GLM-5.2 cloud boost exists — off by default and clearly labeled, because that is the one mode where questions leave the building.)
Everything you put in is kept forever, exactly as it was. You can always see what was true last month — and trust it.
Download a single program. It runs on a laptop or a small server — no install marathon, no database admin.
Open source, no license to renew. Your data, your AI, your machine — nobody else's.
A system of record you can defend in an audit, AI that never sends data outside your perimeter, and a cost line that doesn't grow with usage. Drops into your stack through standard protocols.
Every fact is chained with SHA-256 and per-segment Merkle roots. You can prove your records were never altered — without a blockchain.
Bi-temporal history reconstructs any moment — what was true and what you believed then. The audit superpower for SOX/HIPAA-style review.
Crypto-erasure destroys a record's readability for GDPR while leaving the audit chain intact. Delete data without breaking the trail.
The AI runs locally — sensitive data never leaves your network, and every model answer is itself an auditable, queryable fact.
OIDC/JWT single sign-on (Okta, Azure AD, Auth0, Keycloak), row-level security, and field-level masking — no extra identity glue.
Automated retention policies plus enforced legal hold that blocks deletion of records under investigation.
Automatic failover via leader election: if the primary dies, a replica promotes itself — no operator, no downtime drill.
No per-token AI bills and no database license. Offload high-volume transactions to free Postgres — adopting Centauri never means buying another license.
Speaks the PostgreSQL wire protocol (Tableau, Power BI, JDBC connect directly) and uses S3-compatible storage for a cheap, verifiable cold tier.
A document store, vector DB, time-series, graph, and tamper-proof ledger in one zero-dependency Go binary, with a query language where time, cause, and trust are syntax — and a local-AI stack wired in.
Pure Go standard library — the require block in go.mod is empty. No supply chain, no runtime, one static binary.
Query language with AS OF / AS KNOWN AT, causal WHY/EFFECTS/MATCH, and an NL→CeQL translator for the long tail.
Analytics no other DB ships — SHAPE (persistent homology), CONSISTENCY, CYCLES, DRIFT — plus SNAPSHOT/ROLLBACK/DIFF and PROFILE.
Simple and extended protocol — psql, JDBC and BI tools connect and run read-only SQL that transpiles to CeQL.
ASK = RAG with citations · ENRICH runs models in queries · hybrid BM25+vector SEARCH · vision extraction · centauri desktop / serve -ai auto-provision tiered models up to GLM-4.7-Flash (30B MoE, 200K context).
New facts embed themselves on write; user ratings re-rank retrieval — deterministic, auditable "learning," no weights touched.
An MCP server makes agents first-class clients (grounded memory, no token cost to remember), plus a zero-dependency Python SDK.
Lazy disk-backed index, sharding, group commit, compressed cold segments with zone-map skipping, and an S3 backend with Merkle verification.
Per-line hash chain, online verification, durable CDC slots, log shipping, snapshot/clone, and HA failover — all in the box.
Embedded dashboard, AI Studio IDE, a full CeQL textbook, autocomplete catalog, Airbyte connectors, and the Genesis Engine that builds a DB from a conversation.
Every query below runs verbatim on today's engine (try them on the demo seed: centauri seed).
-- time travel: what was true then — and what we *believed* then FACTS OF item:100001/store:4001 AS OF '2026-03-15' AS KNOWN AT '2026-03-01' -- why did it change? walk the causal chain FACTS OF item:100001/store:4001 WHY DEPTH 3 -- aggregate with GROUP BY / HAVING FACTS facet, COUNT(*) OF item:*/store:4001 GROUP BY facet HAVING COUNT(*) > 5 -- ranked keyword search — and hybrid keyword + vector SEARCH 'late markdown' OF item:* SEARCH 'markdown' OF item:* SIMILAR TO 0193fa2e-77c1 ALPHA 0.5 -- ask your own data: local RAG, answers cite their source facts ASK 'does it scale?' -- the shape of the data: clusters, loops, periodicity SHAPE OF item:*/store:4001 ON price_cents -- run a stored CePL procedure, with a full step-by-step trace RUN duty_estimate WITH item='100001', units=3
Full language reference: The CeQL Book — every statement, with a SQL/Mongo/Cypher Rosetta Stone.
Every capability, grouped. (Honest by policy — see "where it fits" below.)
Centauri is honest about its lane; the right tool earns trust.
One binary · no cloud · no license · nothing leaves your building.