Agentic AI Orchestration

Deploy AI agents on-premises.
Data stays local.

Natural-language database agents, deployed to client networks.

See the architecture Contact
0
data stored off-prem
100%
of sql ast-validated
5
database engines
3-way
llm failover
Why

You have all this data.
Are you using it?

[ask]
No developer needed

Ask in plain English. Get real answers from live data — no SQL, no ticket, no waiting.

[live]
Not a chatbot

ChatGPT can't see your database. This is wired into it — live schema, validated queries, real results.

[auto]
Answers become actions

Monitors watch your metrics around the clock and act the moment something changes.

01

Ask the database anything

English in. Validated SQL and answers out.

→ how many orders last month, by region? generated · validated · executed read-only · narrated
02

Agents that keep watch

Monitors that trigger actions — behind circuit breakers.

→ if orders drop below 10/hr, alert ops compiled to SQL · scheduled · rate-limited
Architecture

Management and Data, permanently separated.

HUB · CONTROL PLANE
runs centrally — sees metadata only
  • Fleet management
  • Agent personas
  • Usage analytics
  • Onboarding wizard
heartbeat · config · usage
query data: never
NODE · DATA PLANE
runs on client LAN — data stays here
  • NL→SQL engine
  • 5 database connectors
  • Scheduled monitors
  • Local vector memory
Reliability

The model proposes. Deterministic code decides.

01Agent loopone model, tools, bounded turnsllm
02Schema retrievalpgvector similarity searchno llm
03Generate SQLdialect-awarellm
04ValidateAST parse — blocks writes & exfildeterministic
05Executeread-only · 30s timeout · 10K row capsandboxed
06Auditevery query and token loggedalways

Fail-safe by default. Errors feed back to the model. Tool loops are capped at 8 turns.

Security

Secure by architecture, not by promise.

[egress]
DLP boundary

One choke point. No opt-outs.

[crypto]
Encrypted credentials

Fernet + key rotation. TLS everywhere.

[sql]
Read-only by default

AST-validated. Mutations gated.

[inject]
Untouchable validator

Client config never reaches the safety layer.

[rbac]
RBAC + audit

Role-scoped. Every action logged.

[bound]
Contained autonomy

Autonomous ≠ unbounded.

Interface

Mission control, both sides.

Node query chat: natural-language question answered from a live database, with the agent tool loop shown alongside
node · query chat + agent loop
Hub mission control dashboard: fleet health, query volume, LLM cost
hub · mission control
Node activity log: full audit trail of tool calls, SQL and model inferences with latencies
node · audit trail
Hub guided client onboarding wizard
hub · guided client onboarding
Agent node fleet view with per-node CPU, RAM and heartbeat
hub · agent node fleet
Usage analytics: query volume, token consumption, cost allocation
hub · usage & cost analytics
Stack

Boring, proven pieces.

python / fastapi / asyncio next.js / react postgresql + pgvector redis anthropic · openai · gemini mcp sqlglot docker / caddy