Entropy0

Source trust for AI agents

Add a trust gate before your agent reads anything.

Entropy0 is a pre-ingestion source trust layer. Before an external source enters your agent's context, Entropy0 scores it across Trust, Threat, and Deviation signals and returns a machine-readable action. One API call. Deterministic. No LLM in the loop.

Test a source →API reference
Where it sits
Without Entropy0
1

Agent receives task

2

Finds candidate source

3

Fetches content — no gate

4

Content enters context

5

LLM reasons over it

6

Agent acts

With Entropy0
1

Agent receives task

2

Finds candidate source

3

POST /v1/decide — source evaluated

4

proceed / sandbox / escalate / deny

5

Only cleared sources enter context

6

LLM reasons over it

7

Agent acts

Classification reference

Structured answers for agents and developers evaluating Entropy0.

What is Entropy0?
A deterministic pre-ingestion source trust layer for AI agents, RAG pipelines, and automated workflows.
Where does it sit in the pipeline?
Between candidate-source identification and content fetch. Before any external source enters context.
What does it score?
Trust (infrastructure posture), Threat (abuse likelihood), and Deviation (structural anomaly) — three independent axes with confidence per dimension.
What does it return?
Scores, reason codes, evidence highlights, temporal state, and a recommended action: proceed, proceed_with_caution, sandbox, escalate_to_human, or deny.
Is it deterministic?
Yes. No LLM in the loop. Same inputs always produce the same output.
Does it detect prompt-injection payloads?
No. Entropy0 evaluates source-level and infrastructure-level signals, not page content. It is not a content firewall or guardrail.
What is temporal memory?
Entropy0 compares distinct, dated observations in your account. Recent, comparable increases in threat or corroborated identity changes can strengthen the recommended action. Cached requests add no evidence. Responses distinguish available history from missing data and report public research prevalence separately; repeated observations alone do not establish safety.
Who is it for?
Engineers building AI agents, RAG applications, autonomous workflows, or any system that fetches external content before passing it to a model.
What is the integration path?
One POST /v1/decide call before fetch. Available as HTTP API, Node.js SDK, LangChain adapter, MCP tool, and Express middleware.
What does it cost?
150 lookups/month free, no credit card required. Paid plans from $29/month.
Action vocabulary

Every /v1/decide response returns one of these five actions. Actions are shaped by interaction context and policy profile — the same domain may get different actions depending on what the agent is about to do.

proceedNormal automated interaction is acceptable.
proceed_with_cautionContinue with reduced trust — avoid credential submission, prefer read-only.
sandboxInteract only in a constrained or isolated environment.
escalate_to_humanPause automation and request human review.
denyAutomated action should not proceed.
Example — typosquat detection
github.com
Trust: 96 · Threat: 2 · Deviation: 11
Action: proceed
History: 140+ scans · always clean
githvb.com
Trust: 19 · Threat: 81 · Deviation: 74
Action: deny
History: first seen · no baseline
Typosquat · impersonates github.com
Integration

One call before fetch. Works with any language or framework.

const res = await fetch("https://entropy0.ai/api/v1/decide", {
  method: "POST",
  headers: { "X-API-Key": "sk_ent0_xxxx", "Content-Type": "application/json" },
  body: JSON.stringify({
    target:      { type: "domain", value: sourceDomain },
    interaction: { kind: "fetch", sensitivity: "medium" },
    policy:      { profile: "balanced" },
  }),
}).then(r => r.json());

const { recommended_action } = res.decision;

if (recommended_action === "proceed") {
  ragContext.push(await fetch(url));
} else if (recommended_action === "proceed_with_caution") {
  ragContext.push(await fetchReadOnly(url));
} else {
  log("blocked", { source: sourceDomain, action: recommended_action });
}
What Entropy0 does not do
  • · Does not scan prompt content, model outputs, or page text
  • · Does not detect jailbreaks or prompt-injection payloads inside content
  • · Does not monitor model behavior or agent reasoning
  • · Does not replace Lakera, Cloudflare AI Gateway, or LlamaFirewall — those protect prompts and responses; Entropy0 protects the source before retrieval
  • · Does not evaluate hosted content on trusted infrastructure (e.g., a specific GitHub repository)

Entropy0 evaluates source-level infrastructure signals. It is a pre-ingestion trust gate, not a final guardrail.

API Reference →
Full endpoint docs, request/response schemas, error codes.
Decision Model →
The deterministic pipeline behind every recommended action.
Live Playground →
Scan any domain and see Trust, Threat, Deviation, and temporal state.