Security posture
Read each agent's findings and score, and the fleet coverage report, and know exactly what each number counts.
- Read an agent's posture findings
- Read the fleet coverage report
- Tell a watched control from an enforced one
- Know which writes count as unapproved
Governance › Posture lists what is wrong with how each agent is set up. Every finding is read from stored configuration. No model is called and nothing is dialled, so the same setup always gives the same findings, and a finding clears as soon as its cause is fixed.
What one agent's report holds
GET /api/agents/{key}/posture returns the findings, each with a severity, a sentence and the page where it is fixed, plus:
| Field | Meaning |
|---|---|
score | 100 minus the weight of each finding (critical 25, high 15, medium 7, low 3), never below 0 |
penalty | The same sum without the floor, so two agents at 0 can be told apart |
grade | A to F, from the score |
counts | Findings per severity |
What the fleet report counts
GET /api/governance/coverage gives, for each control, how many of the agents it applies to have it, and which agents are the gaps.
| Control | An agent is covered when |
|---|---|
cedar | an enabled Cedar policy names it, its blueprint or every agent as the principal. A log-only policy counts: it watches, and refuses nothing |
guardrails | every required guardrail kind reaches it in enforce mode |
tools | every server it reaches is authenticated, encrypted, reviewed, and its writes are held (see below) |
kill_switch | an enabled detector watches it, its own or a tenant default. A shadow or warn detector counts: it alerts, and contains nothing |
owner, budget, recording, exposure | it has an owner, a spend cap, recording on, and a card that is not published unscreened |
Two controls can therefore read "covered" while only watching: cedar with a log-only policy and kill_switch with a shadow or warn detector. That is by design. Start every control in its watching mode and promote it on evidence.
Guardrails are stricter. A monitor-mode rule does not close the gap, since it changes nothing. The finding stays high, names the rule that is already watching, and links to it. The coverage row lists those agents under monitoring, so you can see which gaps are one promotion away.
Which writes count as unapproved
A write tool on a server is reported as going out unapproved (mcp_unguarded_writes) unless one of these stands in front of it:
- the server's Agent writes posture, or a per-tool posture, holds or denies it
- a data policy is attached to the server
- the person the agent acts for confirms the call first. This is how every SAP connector works, and every write through the SAP Integration Suite MCP Gateway route: the call waits on a Confirm card and is refused when there is nobody to confirm it
- on a remote server with no Agent writes setting, the tool is pinned as destructive, which holds it for approval.
A stdio server is never covered by a confirmation. A remote server whose tools were never listed counts every tool as a write.
Publishing an agent's card is refused while it has unapproved writes. Because SAP writes the person confirms no longer count, an agent whose only writes are SAP ones can now be published.