Workspace isolation
Candidate content must belong to the current task workspace.
← All capabilities
Governance / Runtime Control Plane
Governance is not an after-the-fact log.It controls every execution before it runs.
Before information enters context, a tool runs, or business data is written, NexusClaw determines who is acting, what they may read, the action risk, and whether a person must approve.
After execution or blocking, the same chain preserves trace, tool calls, approvals, risk, cost, and outcome evidence. Governance is part of runtime, not a side report.
Product UI · Guardrail rulesNexusClaw Product UI
01 / Permission-aware Context
Unauthorized content is not hidden after retrieval. It never enters Agent context.
Knowledge retrieval checks workspace, role, object-read permission, and content access level for every candidate. If any condition fails, the candidate is denied by default and cannot influence later tool use.
Authorization subject
Independent employees use their role; assistants inherit the real caller’s role.
Object read access
Business objects and related records must be readable by that subject.
Content access level
Private, organization-node, and global knowledge are evaluated individually.
Product UI · Knowledge access levelsNexusClaw Product UI
02 / Guardrails & Risk Levels
The same action can be allowed, reviewed, confirmed, approved, or blocked by risk
Guardrails turn “should not” prompt advice into system policy. Rules can target objects, operations, sensitive fields, tools, and learning usage with deterministic priority.
Allow
Low-risk reads or verified actions run and remain traced.
Review
Extra checks or result review without necessarily pausing the flow.
Confirm
Explicit confirmation before a sensitive action runs.
Approve
Pause execution and route the decision to an accountable person.
Block
Deny the action with no bypass path.
Policy ownership
Guardrail release and rollback have registered owners. An Agent cannot disable, rewrite, or bypass policy at runtime.
03 / Human Approval
Human takeover happens before the sensitive action runs
The approval center gives the approver the object, execution ID, current step, submission time, and recent action. The person reviews context and approves or rejects before execution.
01
Explicit object
Know which business record the action affects.
02
Explicit step
See where execution paused and what it intends to do.
03
Explicit ownership
The real approver identity enters the audit chain.
Product UI · Sensitive Agent action approvalNexusClaw Product UI
04 / Trace, Cost & Audit
Start from a business outcome and reconstruct every tool call, cost, and control decision
Execution detail preserves input, output, status, duration, model and token usage, tool-call chain, and ReAct step timeline. Audit connects risk, object, action, approval, PII, and learning-use state.
Execution trace
Trigger source, step timeline, and final output.
Tool calls
Which tools ran, with parameters and result state.
Model and cost
Model, tokens, duration, and spend by step.
Approval and blocking
Which policy matched, who decided, and final state.
Product UI · Execution detail and tool chainNexusClaw Product UI
Audit turns runtime control into searchable evidence
Filter by execution ID, trace, object, risk, action, and time to see how reads, writes, creates, or blocks happened and whether evidence entered learning.
Product UI · Audit eventsNexusClaw Product UI
Customer Evaluation Checklist
When validating governance, actively try to make the system refuse you
01
Request sensitive knowledge as an unauthorized user
Confirm it never enters context instead of being masked after generation.
02
Trigger L3 and L4 actions
Confirm L3 pauses for approval and L4 blocks without a bypass.
03
Replay execution from a business result
Inspect input, tools, model, cost, approval, and outcome as one chain.
04
Return from an audit event to the real object
Verify linkage to execution, business object, owner, and learning-use state.
Current capability boundary
Governance depends on explicit identity, roles, object access, guardrails, and approval configuration. It does not define your compliance policy for you or imply undeclared certifications; it enforces configured boundaries and preserves evidence.
Bring one high-risk business action to the demo
We will test permission context, guardrail matching, human approval, execution or blocking, then replay the complete audit evidence from the outcome.