You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Part of epic #25 (E7). Milestone: Phase L (day-0 adoption wedge). Source: docs/EPICS.md / docs/SITH-NOTION.md.
F7.1 — MCP read tools
What it is. Read tools (fleet.inventory, fleet.health, fleet.correlate, fleet.cve-search) carrying readOnlyHint: true, hitting the fleet model, scoped to the caller's
workspace.
How it works.
An external agent calls a read tool over MCP.
The MCP layer is a thin client onto the same PEP; the read is tenant-scoped to the caller's
workspace.
The fleet model (E2) answers, including cross-cluster correlation and CVE search.
The read is audited like any other (E6); no gate beyond tenant scope is needed for reads.
sequenceDiagram
autonumber
participant AG as External agent (MCP client)
participant MCP as Sith MCP server
participant PEP as PEP (tenant scope)
participant FM as Fleet model (E2)
AG->>MCP: call fleet.correlate (readOnlyHint: true)
MCP->>PEP: same path as UI — resolve workspace scope
PEP->>FM: query within caller's workspace
FM-->>PEP: cross-cluster answer (stale flagged)
PEP-->>MCP: result (audited)
MCP-->>AG: result
Loading
Acceptance criteria.
Read tools return workspace-scoped fleet answers, including correlation and CVE search.
Reads are audited and carry readOnlyHint.
Key risk / guardrail. A read tool that ignored tenant scope would leak cross-tenant fleet
data. Guardrail: MCP reads go through the same PEP tenant scoping and DB RLS as the UI — the MCP
layer has no privileged data path.
Spec & slice. Slice 6 of docs/BUILD-SEQUENCE.md (PR #40). AC: the four read tools carry readOnlyHint:true, are workspace-scoped and audited; MCP reads use the same scope-resolution path as CLI/UI (no privileged data path — asserted by test); sith serve --mcpbinds loopback only in local mode. Adopts MCP-2026 auth hardening from E7 #25 (audience-bound tokens, enforce-at-execution).
Part of epic #25 (E7). Milestone: Phase L (day-0 adoption wedge). Source:
docs/EPICS.md/docs/SITH-NOTION.md.F7.1 — MCP read tools
What it is. Read tools (
fleet.inventory,fleet.health,fleet.correlate,fleet.cve-search) carryingreadOnlyHint: true, hitting the fleet model, scoped to the caller'sworkspace.
How it works.
workspace.
sequenceDiagram autonumber participant AG as External agent (MCP client) participant MCP as Sith MCP server participant PEP as PEP (tenant scope) participant FM as Fleet model (E2) AG->>MCP: call fleet.correlate (readOnlyHint: true) MCP->>PEP: same path as UI — resolve workspace scope PEP->>FM: query within caller's workspace FM-->>PEP: cross-cluster answer (stale flagged) PEP-->>MCP: result (audited) MCP-->>AG: resultAcceptance criteria.
readOnlyHint.Key risk / guardrail. A read tool that ignored tenant scope would leak cross-tenant fleet
data. Guardrail: MCP reads go through the same PEP tenant scoping and DB RLS as the UI — the MCP
layer has no privileged data path.
Spec & slice. Slice 6 of
docs/BUILD-SEQUENCE.md(PR #40). AC: the four read tools carryreadOnlyHint:true, are workspace-scoped and audited; MCP reads use the same scope-resolution path as CLI/UI (no privileged data path — asserted by test);sith serve --mcpbinds loopback only in local mode. Adopts MCP-2026 auth hardening from E7 #25 (audience-bound tokens, enforce-at-execution).