Get Started

MCP Prompts

AgenticReality exposes four MCP prompts. Prompts are reusable message templates that LLM clients can invoke to get context-aware system prompts or structured queries.

Prompt List

Name	Description
`reality_context`	Inject full operational context as a system prompt
`reality_pre_action`	Get a stakes-aware pre-action safety check
`reality_hallucination_check`	Check a set of claims against grounded reality
`reality_status_report`	Generate a structured status report for the current context

Prompt Schemas

reality_context

Generates a system prompt that injects the current operational context into the LLM's understanding. Useful as a prefix to any conversation where the LLM needs to be aware of WHERE it is operating.

Arguments: none

Sample output:

You are operating in a production environment (us-east-1, cloud substrate).
Current incarnation: inc-a7f3b2 (started 4h ago, restart #3 of this deployment).
Resources: memory 73%, CPU 45%, disk 31% — all within normal operating range.
Reality: 3 ground-truth anchors verified, coherence score 0.98.
Stakes: HIGH. This environment handles financial transactions for 50,000 active users.
Blast radius of an error: 50,000 users. Apply proportional caution to irreversible actions.
Topology: upstream auth-service (healthy, 4ms) and database-primary (healthy, 12ms).

reality_pre_action

Generates a stakes-aware safety assessment before a proposed action. Intended to be called before any consequential operation.

Arguments:

Argument	Required	Description
`action`	yes	Description of the proposed action
`reversible`	no	Whether the action is reversible (`true`/`false`)

Sample output (for action: "deploy new API version to production"):

Pre-action safety check for: "deploy new API version to production"

Stakes level: CRITICAL (production, peak hours, 50,000 active users)
Reversible: NO (deployment to production)
Blast radius: 50,000 users across 12 dependent services

Risk factors:
- auth-service is a single point of failure in the topology
- Current time: Friday 4:47 PM (peak traffic window)
- Memory pressure: 73% (elevated)

Recommendation: DEFER. Consider:
1. Deploy to staging first and run smoke tests
2. Wait for the 2 AM Saturday maintenance window
3. Ensure rollback plan is ready before proceeding

reality_hallucination_check

Checks a list of claims against grounded reality anchors and returns a risk assessment for each.

Arguments:

Argument	Required	Description
`claims`	yes	Comma-separated list of claims to check

Sample output:

Hallucination check results:

✓ "database is healthy" — VERIFIED (anchor match, confidence 0.94, evidence: response 12ms)
✓ "feature_x is enabled" — VERIFIED (anchor match, confidence 0.91, evidence: config v2.3)
✗ "API handles 50,000 RPS" — UNGROUNDED (no matching anchors, risk: HIGH)
? "memory usage is acceptable" — PARTIAL (anchor: mem=73%, claim is vague, risk: LOW)

reality_status_report

Generates a structured status report for the current operational context. Useful for logging, monitoring, or sharing context with human operators.

Arguments:

Argument	Required	Description
`format`	no	`text` (default) or `json`

Sample output:

AgenticReality Status Report — 2025-03-01T14:22:00Z

Deployment
  Incarnation: inc-a7f3b2 (restart #3, running 4h 12m)
  Tier: production | Region: us-east-1 | Substrate: cloud
  Purpose: api-server

Resources
  Memory: 73% used (pressure: normal)
  CPU: 45% used (pressure: normal)
  Disk: 31% used (pressure: none)

Reality
  Ground truth anchors: 3
  Coherence score: 0.98 (no violations)
  Freshness: all layers fresh

Stakes
  Level: HIGH | Blast radius: 50,000 users

Topology
  Upstream: auth-service (healthy), database-primary (healthy)
  Siblings: api-server-1, api-server-2
  Observers: prometheus

Client Usage

Prompts are invoked using the MCP prompts/get method:

{
  "method": "prompts/get",
  "params": {
    "name": "reality_pre_action",
    "arguments": {
      "action": "delete all cached user sessions",
      "reversible": "false"
    }
  }
}

Most MCP clients allow you to trigger prompts through slash commands or the prompt palette.