AgentGate Governance

AgentGate Governance provides the control plane for agent operations. It includes a policy engine, compliance templates, MCP observability, drift detection, and a secrets vault.

Policy Engine

The policy engine evaluates every agent action against a set of rules before execution. Policies are defined as JSON documents and stored per-tenant.

Rule Types

Evaluation Order

Policies are evaluated in this order:

1. Deny rules -- if any deny rule matches, the action is blocked immediately.
2. Approval rules -- if an approval rule matches, the action is queued for human review.
3. Budget and rate limits -- checked against current counters.
4. Time windows -- checked against the current time in the tenant's timezone.
5. Allow rules -- if no allow rule matches, the action is blocked by default (deny-by-default).

All policy evaluations are recorded in the audit trail.

Compliance Templates

Pre-built rule sets for common regulatory frameworks. Each template installs a set of policies that enforce the framework's requirements.

Available Templates

Templates can be customized after installation. Each policy within a template can be individually enabled, disabled, or modified.

Applying a Template

Templates are applied through the dashboard at /dashboard/policies or via the API:

curl -X POST https://api.codespar.dev/v1/policies/templates/sox \
  -H "Authorization: Bearer $API_KEY" \
  -H "x-org-id: $ORG_ID"

MCP Observability

The MCP (Model Context Protocol) observability layer monitors all agent-tool interactions and provides metrics, anomaly detection, and cost tracking.

Tool Metrics

For every MCP tool call, the system records:

• Call count -- total invocations per tool, per agent, per time window.
• Success rate -- percentage of calls that completed without errors.
• Latency -- average and P95 response times.
• Cost -- estimated cost per call based on token usage and API pricing.
• Trend -- directional indicator (stable, up, down) based on the last 24 hours.

Hallucination Detection

The system detects three types of hallucinations:

1. Unknown tool -- the agent attempted to call a tool that does not exist in its MCP schema.
2. Invalid params -- the agent provided parameters that do not match the tool's schema.
3. Type mismatch -- the agent provided values of the wrong type for a tool parameter.

Each hallucination event is logged with the agent ID, tool name, reason, and timestamp. Repeated hallucinations trigger an alert to the tenant's notification channel.

Anomaly Detection

Statistical anomaly detection identifies unusual patterns in tool usage:

• Spike detection -- sudden increase in call volume for a specific tool.
• Error rate deviation -- success rate drops below the 7-day rolling average by more than 2 standard deviations.
• Cost anomalies -- per-agent cost exceeds the 30-day rolling average by more than 50%.

Cost Reports

Cost reports break down spending by agent, by tool, and by time period. Reports are available at /dashboard/observability and can be exported as CSV.

Drift Detection

Drift detection monitors the APIs that MCP tool definitions depend on. When an external API changes its schema, endpoints, or behavior, the system detects the mismatch and alerts the tenant.

How it works

1. The system periodically probes each registered API endpoint with a lightweight health check.
2. Response schemas are compared against the stored MCP tool definitions.
3. If a field is added, removed, or its type changes, a drift event is created.
4. The drift event includes the affected tool, the specific field, the expected type, and the actual type.

Drift events are surfaced in the dashboard and sent as notifications to the configured channel.

Secrets Vault

The secrets vault provides AES-256-GCM encrypted, per-tenant credential storage for API keys, tokens, and other sensitive values that agents need to access external services.

Features

• Per-tenant isolation -- each tenant's secrets are encrypted with a unique key derived from the tenant ID and a master key.
• Access control -- secrets are scoped to specific agents and tools. An agent can only access secrets it has been explicitly granted.
• Rotation support -- secrets can be rotated without downtime. The vault maintains the previous version for a configurable grace period.
• Audit logging -- every secret read, write, and rotation is recorded in the audit trail.
• No plaintext storage -- secrets are encrypted at rest and decrypted only in memory at the moment of use.

Dashboard

The secrets vault is managed through the dashboard at /dashboard/secrets. The UI allows creating, viewing (masked), rotating, and deleting secrets. Secret values are never displayed in full after creation.

Dashboard Pages

AgentGate Governance features are managed through three dashboard pages: