Most AI agents today can call APIs.
But very few systems solve the real problem:
how do you safely expose business data to AI agents without giving them raw database access?
That’s what we built in Adeloop.
Introducing: Adeloop Agent Console API
Adeloop can now publish semantic domains as governed APIs for external AI agents and applications.
The flow is simple:
- Connect a warehouse, database, spreadsheet, or file source
- Turn tables into a semantic domain
- Publish selected domains
- Generate an API key
- Connect from ChatGPT, Claude, Cursor, n8n, Zapier, or your backend
The important part:
External AI agents never access raw SQL directly.
Adeloop becomes the governed execution layer between AI and data.
Why This Matters
Most “AI data chat” products are either:
- unsafe SQL generators
- notebook wrappers
- or vector search over metadata
That breaks quickly at scale.
Instead, Adeloop uses:
question
→ semantic routing
→ metric/dimension planning
→ bounded SQL compilation
→ source pushdown execution
→ governed JSON response
This means:
- queries stay close to the warehouse
- millions of rows are not pulled into Python
- agents receive structured JSON
- governance and rate limits stay enforced
The default execution mode is:
semantic_sql_pushdown
Not Python.
Not sandbox compute.
Not “LLM writes random SQL”.
Example
An external agent can ask:
{
"question": "Show top customers by revenue",
"limit": 10
}
Adeloop then:
- selects the semantic domain
- resolves semantic metrics/dimensions
- compiles safe SQL
- executes against Postgres/MySQL/Snowflake/etc
- returns answer + JSON + execution metadata
Example response:
{
"answer": "Top result is Acme with total_revenue = 124500",
"execution": {
"mode": "semantic_sql_pushdown",
"engine": "postgresql",
"sandboxUsed": false
}
}
MCP + OpenAPI Support
We also added:
- MCP-compatible JSON-RPC endpoint
- OpenAPI 3.1 action schema
- API key scopes
- usage logs
- semantic metadata endpoints
- deterministic domain routing
So tools like:
- adeloopchat
- Claude tools
- Cursor
- n8n
- ChatGPT Actions
can consume governed business data without direct warehouse access.
One Important Architecture Decision
We intentionally did NOT add E2B/sandbox execution into the main API path.
Why?
Because most business questions are:
- aggregations
- grouped metrics
- dashboards
- top-N queries
- filters
- time-series analytics
Those should execute through SQL pushdown near the data source.
Python notebooks and sandbox compute belong later as async premium analysis jobs for:
- forecasting
- anomaly detection
- ML
- simulations
- notebook/report generation
Normal analytics APIs should stay fast, deterministic, and scalable.
The Bigger Goal
I think AI agents will need something equivalent to:
a semantic execution layer for enterprise data
Not just chat over databases.
Something that handles:
- governance
- semantic metrics
- execution planning
- safe query compilation
- federation
- caching
- observability
- API contracts for agents
That’s the direction we’re building toward with Adeloop.
Would love feedback from people building:
- AI agents
- semantic layers
- MCP tooling
- data infrastructure
- analytics engineering systems