AI Graph

AI Graph is the governed view of how everything in the mesh connects — actors, MCP servers, tools, resources, data assets, and the policies protecting them.

The backend model is a directed edge: a (src_type, src_id) → (dst_type, dst_id, edge_type) row, written idempotently so re-running discovery doesn't duplicate the graph. Building the graph for display collects every distinct node referenced by an edge, then resolves a human-readable name for each from the matching catalog table — MCP server, tool, resource, model version, agent, application, dataset, table, column, prompt template, or role — so the page never shows a bare UUID.

The page itself has four tabs. By MCP and By Role are flat card lists; By Actor groups agents into confidence bands — observed, approved, curated, blocked — with observed shown first since those are the actors still waiting on operator review. The fourth tab, AI Estate, is the unified graph: every governance path laid out across five columns (Identity → MCP Server → Capability → Data Asset → Column), with each hop colored by the real policy decision that applies there — deny, conditional, allow, or unprotected — derived from actual policy bindings on the subject and target, not a static label.

Clicking a node in the unified graph isolates every path running through it and dims the rest; tables and columns stay collapsed by default and only appear on click (or automatically when a parent has ten or fewer children), so a large MCP's fan-out doesn't overwhelm the canvas. Every node opens a detail panel with its full hop-by-hop chain.

That detail panel is also where the graph turns into action: a Create Policy button scoped to whatever was clicked — MCP-wide, one tool, one table, one column, one role, or one agent — deep-links straight into the Policy Builder pre-filled with that scope. When a path reaches a sensitivity-tagged node over a hop with no policy at all, a dedicated "Protect …" callout surfaces automatically, so the highest-risk gap on a path is the one an operator is nudged toward first.