Weekend-ready agent grounding

MCP quickstart

• Provision tenant credentials through request access.
• Configure MCP with environment variables only.
• Run retrieval + citation flows before scaling automation.

{
  "mcpServers": {
    "vaultcrux": {
      "command": "vaultcrux-mcp",
      "env": {
        "VAULTCRUX_API_URL": "https://vaultcrux.com",
        "VAULTCRUX_API_KEY": "$VAULTCRUX_API_KEY",
        "VAULTCRUX_TENANT_ID": "$VAULTCRUX_TENANT_ID"
      }
    }
  }
}

Tool catalogue

30 MCP tools across five groups. All tools use standard MCP JSON-RPC.

Group A — Knowledge & Memory

• Ingest: shipped.
• Retrieve with citations: shipped.
• Usage Unit visibility: shipped.
• query_memory — semantic search across memory records with optional team scope.
• list_topics — list topics in the knowledge base.
• get_versioned_snapshot — point-in-time knowledge snapshot.
• get_audit_trail — change history for a topic.
• check_claim — verify a claim against stored knowledge with confidence scoring.
• get_freshness_report — staleness flags per topic.
• get_contradictions — cross-platform divergence alerts.
• get_session_context — session continuity payload with journal, stale pins, watch alerts, and credit balance.
• reconstruct_knowledge_state — temporal reconstruction via CoreCrux — what did the system know at an exact moment?

Group B — Decisions & Causality

• get_decision_context — full decision record: what was known, confidence, tool called, outcome.
• get_causal_chain — navigate the linked graph of decisions that produced an artefact.
• get_correction_chain — walk the supersession graph — what went wrong and what corrected it.
• get_decisions_on_stale_context — decisions made on evidence that has since been superseded.
• record_decision_context — capture a human decision with its full reasoning context.

Group C — Constraints & Verification

• declare_constraint — encode senior judgment as a machine-checkable boundary.
• update_constraint — modify a constraint via append-only supersession.
• get_constraints — list active constraints, optionally filtered.
• check_constraints — check whether a proposed action intersects with any active constraints.
• verify_before_acting — composite pre-flight gate: Shield + Engine + Watch + Constraints → single go/no-go.
• suggest_constraint — agents propose constraints for human review without overstepping authority.

Group D — Monitoring & Health

• get_pressure_status — current pressure level and active signals for an artefact or tenant.
• get_active_alerts — Watch alerts since a given timestamp, filtered by severity.
• get_signals_feed — mechanical status updates on knowledge changes.

Group E — Agent Effectiveness

• assess_coverage — coverage gap analysis: what don’t I know relative to this task?
• get_relevant_context — task-scoped context briefing prioritised by risk-if-missed, within a token budget.
• escalate_with_context — contextual escalation that packages full reasoning state for human review.
• checkpoint_decision_state — receipted snapshot of active decision state for session resumption.
• get_platform_capabilities — machine-queryable manifest of available tools, trust tiers, and credit costs.

External proof surfaces

• Decision Chain Proof Pages — public URLs rendering the full causal chain for any receipted decision.
• Cross-platform receipt anchors — portable cryptographic verification beyond the platform boundary.
• Capability discovery — get_platform_capabilities returns structured service metadata for agent-to-service evaluation.

Modes

• Light: fast iteration.
• Verified: stronger checks for sensitive outputs.
• Audit: highest traceability for high-stakes reviews.

No surprises

Budgets, Usage Units, and explicit failures are surfaced directly so automation can fail safe.