System · Open source

syntropAI

One server per SDK, not one tool per service.

1 coreevery SDK plugs into, reaching all its services
~175 linesis the whole core, each SDK adds about a page
Any SDKAWS, Oracle, Azure, Google Cloud, market data and compliance
The problem

When AI agents learned to use tools, most servers wrapped cloud APIs one tool per service. One provider has hundreds of services, every provider needs its own set, and a new service waits until someone writes its tool.

syntropAI gives the agent the whole SDK instead. The agent writes the code, and every line is checked before it runs.

The insight

Clouds, markets and project tools look nothing alike. Underneath, they share one shape.

CloudFinanceProjects
ResourceVM, storageStock, bondIssue, story
CollectionResource groupPortfolioWorkflow
OperationProvision, scaleBuy, sellCreate, update
ConstraintCost, region, complianceRisk, exposureBudget, timeline
How it works
1AskSay what you want in plain language.
2WriteThe agent writes code against the provider’s own SDK.
3CheckThe code is read as a syntax tree first. Only allowed modules, no dangerous calls, an answer required.
4RunOnly then does it run, and the result comes back.
Under the hood
boto3OCIAzureGooglefinvizOSCAL

Each SDK plugs in with its login and a namespace, about a page of code.

→
syntropaibox
Being addedDescribeReads the SDK’s own catalogue: services, methods and exact parameters.
RunParse into a syntax tree, check imports and calls, run with the session, return JSON.
→
AI agentThrough MCP. It looks up the SDK, writes the code, and gets the result.

Many SDKs, one core, the same two abilities for each.

What was built
Built with
PythonMCPPython ASTAST guardrailsboto3OCI SDKAzure SDKGoogle API clientfinvizfinance ↗NIST OSCAL ↗compliance-trestle ↗PydanticClaude Code