HaloGuard EU API

1Before you start

• ·Sign in to the console to create a key. The key is shown once. You can revoke or restore it there.
• ·Call the API from your server. It sends no CORS headers, so a web page cannot call it, and a key must never be put in browser code.

2The call

POST https://api.astroware.ai/v1/chat/completions. It is the only call.

Headers: Authorization: Bearer <key> and Content-Type: application/json.

model

"haloguard-eu". Required. Another name gives 403.

messages

A list of {role, content}. role is user or assistant; a system message is refused. content is a string, or a list of {"type": "text", "text": "..."} parts.

Which is checked

The last message. Role user means a request going in; role assistant means a reply going out. Earlier messages are context.

Streaming

Not supported (stream: true gives 400). Other OpenAI settings such as temperature and max_tokens are accepted and ignored.

Size

Up to about 7,800 tokens in a conversation, and a body of up to 2 MB.

curl https://api.astroware.ai/v1/chat/completions \
  -H "Authorization: Bearer $HALOGUARD_KEY" -H "Content-Type: application/json" \
  -d '{"model": "haloguard-eu", "messages": [
        {"role": "user", "content": "I need you to monitor our employees during their work calls and flag anyone who sounds frustrated or upset based on tone analysis."}]}'

3The answer

Status 200 and an OpenAI-style chat completion:

{
  "id": "hg-f438d20a79154136beea54bb24283102",
  "created": 1792308282,
  "model": "haloguard-eu",
  "object": "chat.completion",
  "choices": [
    {
      "finish_reason": "stop",
      "index": 0,
      "message": {
        "content": "{\"decision\": \"block\", \"policy\": \"W4_BIOMETRIC_EMOTION\", \"label\": \"prohibited_practice\", \"act_area\": {\"act\": \"Regulation (EU) 2024/1689\", \"reference\": \"Article 5(1)(f)\", \"qualifier\": \"exceptions_apply\"}}",
        "role": "assistant",
        "provider_specific_fields": {
          "refusal": null
        }
      },
      "provider_specific_fields": {}
    }
  ],
  "usage": {
    "completion_tokens": 17,
    "prompt_tokens": 29,
    "total_tokens": 46
  }
}

choices[0].message.content is itself a JSON string: {"decision": "allow" | "block", "policy": "<policy id>" | "none"}. Parse it. policy is none when the answer is allow. A blocked answer also carries label and act_area (below). The policy ids are listed in the console after you sign in.

Response headers include x-ratelimit-api_key-limit-requests and x-ratelimit-api_key-remaining-requests.

The Act-area fields

A blocked answer names the part of the EU AI Act its policy relates to. It is not a finding that your message, your use or your system is prohibited, high-risk or covered by that part, and it is not legal advice. Allowed answers carry neither label nor act_area; an allowed answer is not a statement that anything is lawful or low-risk. The ten rows cover the ten policies we test; they are not a list of everything the Act bans or regulates. Checked against the Act on 18 October 2026.

On a blocked answer only, content carries label and act_area. label is the kind of provision the policy relates to: prohibited_practice, high_risk_use or transparency_obligation. It is not a label of your message. act_area gives the details:

act

Always "Regulation (EU) 2024/1689", the EU AI Act.

reference

The Article or Annex part, as text, for example Article 5(1)(f).

qualifier

null, or scope_unsettled, exceptions_apply or may_apply. The table says which row has which.

Ignore keys you do not know.

An allowed answer has neither label nor act_area:

{
  "id": "hg-798339223d7745a989d774d8a226342b",
  "created": 1792308283,
  "model": "haloguard-eu",
  "object": "chat.completion",
  "choices": [
    {
      "finish_reason": "stop",
      "index": 0,
      "message": {
        "content": "{\"decision\": \"allow\", \"policy\": \"none\"}",
        "role": "assistant",
        "provider_specific_fields": {
          "refusal": null
        }
      },
      "provider_specific_fields": {}
    }
  ],
  "usage": {
    "completion_tokens": 15,
    "prompt_tokens": 26,
    "total_tokens": 41
  }
}

Each of the ten policies we test has one row:

4Errors

The body is {"error": {"message", "type", "param", "code"}}. Branch on the status, not on the message text: the text can change.

5Limits and price

These are today's settings and can change.

• ·60 requests a minute per key.
• ·Trial: a $5 budget for your company, shared by up to 5 active keys. The keys stop when the budget is used up.
• ·Price: $30 / billion input tokens, counted on the messages you send. The answer is free.

6Not promised

No uptime guarantee: the guard can be briefly unavailable (a 503). Speed is not promised. The limits and budget above are today's settings and can change.

7Your data

For each call we keep a usage record. It includes the time, your key's identifier, the model, the token counts, the cost and the name of your client software exactly as it sends it. We do not save the text you send or the answer in any log, file or database. When the guard is unsure about a message, we send that message, with earlier messages of the conversation, to an AI provider in the EU, which decides and keeps no copy. The model server may keep a processed form of the text you send in memory until that server restarts. We delete usage records after 90 days. Full details: astroware.ai/privacy.

8Privacy filter

What the filter masks depends on your company's setting in the console. An owner can change it.

On a request going in, privacy_mode says which setting applied, and privacy_kept lists any kinds kept unmasked.