Agents and MCP
Every capability is a tool. What that means for what an agent can and cannot do.
Connecting
/api/mcp. Authentication is a personal access token in an Authorization: Bearer header. See Connect an agent for the ten-line version.
What a token decides
products:read, decisions:write, infra:read and so on. Scopes do two things:
They gate calls. A tool whose scope you do not hold is refused. They gate discovery. tools/listonly returns tools the calling token can actually use, so an agent never sees a capability it cannot reach. Being refused is the fallback; not being offered is the mechanism.
products:read is a genuinely different thing from one that can also write.
What no token can grant
Autonomy
One call to hand over everything
get_product_context returns a product's whole working picture in one call: the why fields, the current stage, open decisions, the board, recent activity, knowledge, links and more - budgeted so it fits a context window, with an opaque revision marker you can pass back to get only what changed since. It is the call to make at the start of a session, instead of fifteen.