Disposable infrastructure for AI agents
An agent that writes and tests code needs somewhere safe to put data. The MCP server lets an agent create a temporary service, use it and let it expire.
Why would an agent need provisioned services?
Agents that run tests, process files or build prototypes need storage, a database or a mail inbox without a human clicking through a console. A time-limited service keeps an agent mistake from becoming a permanent resource.
How does an agent connect?
The msaas-mcp server speaks MCP over stdio and wraps the same REST API as the command line tool. Give it a deploy-scope key to let the agent create and delete, or a read-scope key for inspection only. Credentials are returned only by the create call.
- Tools: list_catalogue, create_service, list_services, get_service, start_service, stop_service, delete_service, get_usage, get_credits and get_account.
- delete_service is marked destructive.
- All inputs are validated; bad input and refusals come back as tool errors.
What are the safety limits?
Services reach each other over a shared private network, so every service requires authentication. Outbound traffic is not filtered, so a service can fetch any URL.
Which services fit?
| Service | Floor | Status |
|---|---|---|
| Key-value store | 64 MB | Validated |
| Analytical SQL | 128 MB | Validated |
| Lakehouse tables | 128 MB | Validated |
| Messaging | 64 MB | Validated |
| Search | 96 MB | Validated |
| Mail inbox | 64 MB | Validated |
Floors are the smallest tested memory, not sizing advice. How we measure. Pricing.
Last updated