Agent tools

Connecting over MCP gives your agent five tools. Your agent decides when to use them, but this is the shape of a good turn: focus on the intent, act, then report what happened.

sioma_focus#

The main one. Give it the user's intent in natural language, and it returns the minimal grounded context for that intent.

What comes back:

• status — serve, ask, or none. See Overview.
• systemText — the context to give your model.
• verified — true only when this answer has been confirmed by a human, an author declaration, or a benchmark answer key. A previous agent saying "that worked" does not make it verified.
• recordId — pass this back to sioma_record_outcome.

Optional inputs: userId sharpens recall for that person, and sessionId enables follow-up suggestions across a conversation.

sioma_list_entities#

The index of your system: every entity name with a one-line shape. This is for navigation, when your agent wants to know what exists. It is not resolution, so reach for sioma_focus when there is an actual task to ground.

sioma_explore_entity#

One entity, expanded: its fields, its declared relations, and the actions that can be called on it. Takes an entity id from sioma_list_entities.

sioma_resolve_path#

The declared relation chain from one entity to another, with the endpoint to call at each hop. Use it when an answer spans several related entities, for example customer to subscription to invoice.

Every hop is a relation your spec declares. If there is no declared path, you get nothing rather than a plausible-looking route.

sioma_record_outcome#

Report what happened after acting on a focused context. Pass the recordId from the focus that produced it.

This is a self-report. It is recorded as usage signal and weakly sharpens recall, but it never counts as verified correctness, and the wire cannot claim otherwise. Report honestly: correct: false is the more useful signal, and a wrong report costs you recall quality without buying anything.

Include inputTokens and outputTokens and your savings show up in Insights.

With rails on#

Rails are off by default. For a workspace with rails enabled, four more tools appear, and they are the agent's entire action surface: sioma_board to start riding a compiled flow, sioma_provide to fill a value the flow is waiting on, sioma_inspect to check where a ride is, and sioma_abort to stop it.

Your agent never supplies a plan. It names a flow, and the layer compiles and runs it from what your spec declares. A write flow pauses for a human approval before any call is made. If you want rails on for your workspace, talk to us.