Governance Policy MCP Server

Policy evaluation, approvals, simulation, and audit evidence for ADK-Rust Enterprise. The gatekeeper that should not be bypassable by downstream MCPs.

Purpose

Let agents and control-plane workflows evaluate policies before risky actions: tool writes, browser submits, payment intents, memory writes, release promotions, credential access, and protocol calls.

Every other MCP server in the ecosystem calls this server before executing high-risk operations. It decides allow, block, review, or pause.

Tools (8)

| Tool | Purpose | Risk Class | |------|---------|------------| | evaluate_policy | Decide allow, block, review, or pause | Read-only | | request_approval | Open a HITL approval task | External write | | list_approval_queue | Show pending approvals by status | Read-only | | resolve_approval | Approve, reject, or request changes | External write | | simulate_policy_pack | Test policy changes against actions | Read-only | | get_policy_decisions | Read recent allow/block/review outcomes | Read-only | | create_policy_exception | Request time-bound policy exception | External write | | export_audit_pack | Export governance evidence for compliance | Read-only |

Risk Domains

| Domain | Default Decision | Rationale | |--------|-----------------|-----------| | tool_write | Allow | Low-risk internal writes | | external_write | Allow | Standard external actions | | financial_action | Review | Requires human approval | | identity_action | Pause | Requires identity verification | | production_deploy | Review | Release governance required | | memory_write | Allow | Standard memory operations | | credential_access | Allow | If scoped correctly | | data_export | Review | Compliance check required |

Example Prompts & Outputs

Evaluate a financial action

Prompt: "Can this agent process a refund?"

Tool: evaluate_policy

{ "action": "process_refund", "actor": "support_agent", "risk_domain": "financial_action" }

Output:

{
  "decision_id": "dec_2e106e4d...",
  "decision": "review",
  "policy_id": "pol_financial_actions",
  "action": "process_refund",
  "actor": "support_agent",
  "risk_domain": "financial_action",
  "reason": "Financial actions require human approval",
  "conditions": [],
  "evaluated_at": "2026-05-23T12:58:26Z"
}

Evaluate a low-risk action

Prompt: "Can agent_1 read configuration?"

{ "action": "read_config", "actor": "agent_1", "risk_domain": "tool_write" }

Output:

{
  "decision": "allow",
  "policy_id": "pol_default",
  "reason": "Action 'read_config' allowed by default policy"
}

Request approval for a blocked action

Prompt: "Request manager approval for the refund"

Tool: request_approval

{
  "action": "process_refund",
  "actor": "support_agent",
  "risk_domain": "financial_action",
  "reason": "Customer requested $89.99 refund",
  "assigned_to": "manager_jane"
}

Output:

{
  "approval_id": "appr_e852cdb0...",
  "action": "process_refund",
  "status": "pending",
  "assigned_to": "manager_jane",
  "expires_at": "2026-05-24T12:58:26Z"
}

Resolve an approval

Prompt: "Approve the refund request"

Tool: resolve_approval

{
  "approval_id": "appr_e852cdb0...",
  "decision": "approved",
  "resolved_by": "manager_jane"
}

Output:

{
  "approval_id": "appr_e852cdb0...",
  "status": "approved",
  "resolved_by": "manager_jane",
  "resolved_at": "2026-05-23T13:15:00Z"
}

Simulate a policy pack

Prompt: "What would happen if these 3 actions were attempted?"

Tool: simulate_policy_pack

{
  "actions": [
    { "action": "send_email", "actor": "agent_1", "risk_domain": "external_write" },
    { "action": "deploy_prod", "actor": "ci_bot", "risk_domain": "production_deploy" },
    { "action": "reset_password", "actor": "admin", "risk_domain": "identity_action" }
  ]
}

Output:

{
  "simulation_id": "sim_7f2a...",
  "total": 3,
  "results": [
    { "action": "send_email", "decision": "allow", "policy_id": "pol_default" },
    { "action": "deploy_prod", "decision": "review", "policy_id": "pol_production_deploy" },
    { "action": "reset_password", "decision": "pause", "policy_id": "pol_identity_security" }
  ]
}

Create a time-bound exception

Prompt: "We need a 4-hour exception for emergency refund batch processing"

Tool: create_policy_exception

{
  "policy_id": "pol_financial_actions",
  "reason": "Emergency refund batch processing",
  "duration_hours": 4
}

Output:

{
  "exception_id": "exc_9d74ab7b...",
  "policy_id": "pol_financial_actions",
  "status": "pending",
  "valid_from": "2026-05-23T12:58:26Z",
  "valid_until": "2026-05-23T16:58:26Z"
}

Export audit pack

Prompt: "Export governance evidence for the compliance audit"

Tool: export_audit_pack

{ "limit": 100 }

Output:

{
  "pack_id": "audit_4f8a...",
  "exported_at": "2026-05-23T13:00:00Z",
  "decisions": 5,
  "approvals": 2,
  "exceptions": 1,
  "policy_decisions": [...],
  "approval_tasks": [...],
  "policy_exceptions": [...]
}

Decision Flow

Action requested → evaluate_policy()
    │
    ├── ALLOW → proceed immediately
    ├── BLOCK → denied, log reason
    ├── REVIEW → request_approval() → wait for resolve_approval()
    └── PAUSE → requires verification before retry

Integration with Other MCPs

| MCP | How It Calls Governance | |-----|------------------------| | Credentials Vault | Before request_runtime_secret and rotate_credential | | Artifact Store | Before write_artifact with sensitive data class | | Session Memory | Before store_memory with PII, delete_memory | | Environment | Before promote_build and scale_worker_pool in production | | ADK-Payments | Before every payment intent execution |

Installation

git clone https://github.com/zavora-ai/mcp-governance-policy
cd mcp-governance-policy
cargo build --release

MCP Client Config

{
  "mcpServers": {
    "governance-policy": {
      "command": "/path/to/mcp-governance-policy"
    }
  }
}

Works with Claude Desktop, Kiro, Codex, Cursor, Windsurf, Antigravity, and Open Code.

Contributing

PRs welcome. Run cargo clippy and cargo fmt before submitting.

Contributors

|
_{James Karanja Maina} | |:---:|

License

Apache-2.0 — see LICENSE for details.

Part of the ADK-Rust Enterprise MCP server ecosystem.

Built with ❤️ by Zavora AI