Give your agent
a next step.

Sentinel does not write code or prose. It reads the evidence and alternatives you provide, then returns a structured decision with model scores. Your agent retains control of execution and verification.

¹ Local RTX 5060 Ti deployment smoke, 12 synthetic requests, September 28, 2026. This is a latency observation, not a quality benchmark or service-level guarantee.

What it does

Three structured interfaces for decisions between agent steps. The outputs below illustrate response formats; they are not measured results.

Why use it

• FastA small 12-request deployment smoke measured a 0.73-second median round trip. Actual latency depends on input length, load and connection.¹
• InspectableCompare scores across the options you supply. These are model estimates, not verified outcomes or universally calibrated confidence. Check consequential choices against evidence.
• Fails fastBounded timeouts and circuit breakers let your agent continue when Sentinel is unavailable. Handle busy or offline responses explicitly in your workflow.

Where it works

• OrrerySentinel runs inside Orrery, where the agent consults it for the quick calls during a run.
• MCP agentsAdd the Sentinel MCP server to any agent that speaks the Model Context Protocol: Claude Code, OpenCode, Cursor, Codex CLI, Gemini CLI or VS Code. The agent gets three tools, sentinel_decide, sentinel_rank and sentinel_status, with descriptions that tell it when to ask.
• Function callingFor your own agent loop, register sentinel_decide as a tool in the OpenAI or Anthropic format and forward the call to the API.
• ShellA small command-line client answers in its exit code, for scripts and shell-driven agents.
• APITwo HTTP endpoints on api.ephemerent.com, with the same key as the Orrery API.

API quickstart

Sentinel uses the API key from your account page, sent as a bearer token. Give it the facts in state (code, a diff, an error message), ask one question, and add options when there are more than two answers.

curl https://api.ephemerent.com/v1/sentinel/decide \
  -H "Authorization: Bearer $ORRERY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "state": "test_parse_date fails: ValueError on \"2026-09-26T10:00Z\". The patch adds a branch for a trailing Z in parse_date.",
    "question": "Will this patch make the failing test pass?"
  }'

# {"answer": "yes", "p_yes": 0.91, "model": "sentinel-1"}

# Python (pip install requests)
import os, requests

r = requests.post(
    "https://api.ephemerent.com/v1/sentinel/decide",
    headers={"Authorization": f"Bearer {os.environ['ORRERY_API_KEY']}"},
    json={
        "state": diff + "\n\n" + test_output,
        "question": "Which of these should the agent do next?",
        "options": {
            "edit": "apply the fix to parse_date",
            "read": "open the caller before changing anything",
            "run": "run the test suite as it is",
        },
    },
    timeout=8,
)
if r.status_code == 503:
    decision = None          # Sentinel is offline: decide without it
else:
    r.raise_for_status()
    decision = r.json()
    # {"answer": "edit", "confidence": 0.84,
    #  "p": {"edit": 0.84, "read": 0.12, "run": 0.04}, "model": "sentinel-1"}

// JavaScript (fetch)
const res = await fetch("https://api.ephemerent.com/v1/sentinel/rank", {
  method: "POST",
  headers: {
    Authorization: `Bearer ${process.env.ORRERY_API_KEY}`,
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    context: failingTestOutput,
    candidates: ["fix the parser", "fix the test"],
    question: "Which fix should we try first?",
  }),
  signal: AbortSignal.timeout(8000),
});
const ranked = res.ok ? (await res.json()).ranked : null; // null: go on without it
// [["fix the parser", 0.78], ["fix the test", 0.22]]

For your own agent loop, this is the tool definition to register. Forward its arguments to /v1/sentinel/decide and return the answer to the model.

{
  "name": "sentinel_decide",
  "description": "Ask Sentinel to compare concrete next actions using the supplied evidence. Returns model scores; verify consequential choices independently.",
  "parameters": {
    "type": "object",
    "properties": {
      "state":    { "type": "string", "description": "The facts the decision depends on: code, a diff, error output." },
      "question": { "type": "string", "description": "One question with a short answer." },
      "options":  { "type": "object", "additionalProperties": { "type": "string" },
                    "description": "Named options, each with a one-line description. Omit for a yes/no question." }
    },
    "required": ["state", "question"]
  }
}

Limits and pricing

Each decision costs a fraction of a cent, metered in the same compute credits as the rest of the API and drawn from the same balance. sentinel-1 is listed in /v1/models, but it is not a chat model: sending it to /chat/completions returns 400. When Sentinel is offline the API answers 503 with a Retry-After header; clients should carry on without it rather than wait. Sentinel requests are never stored or used for training, whatever your data setting: the service keeps counts and an access-log line, never what you sent or what it answered.

An honest note