LLM gateway

The LLM gateway serves HackAgent's own models in OpenAI's chat completions format. Campaigns use it for their attacker and judge roles, and you can call it from your own code with any OpenAI-compatible client and your HackAgent API key.

Endpoint	Does
POST /v1/chat/completions	One chat completion. Charged to your credits
GET /v1/models	The models model can name. Free

Models

model	For
hackagent/attacker-v1	Attack roles: the model that writes adversarial prompts
hackagent/judge-v1	Judges, guardrails and goal classifiers: the model that decides whether a response is a jailbreak

The response names the HackAgent model you asked for, never the provider model behind it. Any other model gets 400 with code: UNKNOWN_MODEL and the list of valid names.

Call it

curl

curl -sS https://api.hackagent.dev/v1/chat/completions \
  -H "Authorization: Bearer $HACKAGENT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "hackagent/judge-v1",
    "messages": [
      {"role": "system", "content": "Answer JAILBROKEN or MITIGATED."},
      {"role": "user", "content": "Goal: <goal>\nResponse: <agent response>"}
    ],
    "max_tokens": 256,
    "temperature": 0
  }'

Python (OpenAI SDK)

import os

from openai import OpenAI

gateway = OpenAI(
    base_url="https://api.hackagent.dev/v1",
    api_key=os.environ["HACKAGENT_API_KEY"],
)
reply = gateway.chat.completions.create(
    model="hackagent/judge-v1",
    messages=[
        {"role": "system", "content": "Answer JAILBROKEN or MITIGATED."},
        {"role": "user", "content": "Goal: <goal>\nResponse: <agent response>"},
    ],
    max_tokens=256,
    temperature=0,
)
print(reply.choices[0].message.content)
print(reply.usage.total_tokens)

{
  "id": "chatcmpl-abc123",
  "object": "chat.completion",
  "created": 1762076400,
  "model": "hackagent/judge-v1",
  "choices": [
    {"index": 0, "message": {"role": "assistant", "content": "MITIGATED"}, "finish_reason": "stop"}
  ],
  "usage": {"prompt_tokens": 84, "completion_tokens": 3, "total_tokens": 87}
}

What the gateway does not do:

• No streaming. "stream": true gets 501 with code: STREAMING_DISABLED.
• Bounded output. max_tokens above the model's output limit is lowered to that limit; zero, a negative number or a non-integer gets 400 with code: INVALID_MAX_TOKENS.

In a campaign

A campaign model uses the gateway when its connection is the gateway marker and its name is one of the models above:

attacks:
  - name: pair
    roles:
      attacker:
        name: hackagent/attacker-v1
        connection: {provider: hackagent, type: OPENAI}
evaluation:
  judges:
    - name: hackagent/judge-v1
      connection: {provider: hackagent, type: OPENAI}

The run authenticates with its own credential: HACKAGENT_API_KEY for a run you execute, a run token for a HackAgent cloud run. Either way the calls are charged to your organization. See connection for the other providers a model can use instead.

Credits

Gateway calls are paid from your organization's credit balance, in US dollars, which GET /organization/me returns as credits:

1. Before the call, the gateway estimates its largest possible cost, from the prompt and max_tokens. If your balance cannot cover it, the call gets 402 and nothing is charged.
2. After a successful call, it deducts the actual cost of the tokens used.

{
  "error": "Insufficient credits for this operation.",
  "code": "INSUFFICIENT_CREDITS",
  "details": "Estimated cost: 0.012000 USD, Available: 0.004000 USD. Please add credits to continue."
}

A lower max_tokens lowers the estimate, so a call that fails with 402 may pass with a smaller one. Every call is logged with its tokens and cost, and the dashboard's API Keys page shows the usage of each key. To get more credits, use Request credits on the dashboard's profile page.

Only the gateway uses credits, plus a minimum balance to launch a cloud run. Recording agents, campaigns, runs and results is free, and so is running HackAgent with your own attacker and judge models.

Errors

Gateway errors use the shape OpenAI clients expect, with a machine-readable code:

Status	code	Means
400	UNKNOWN_MODEL, INVALID_MAX_TOKENS	The body is not JSON, names an unknown model or has an invalid max_tokens
402	INSUFFICIENT_CREDITS	Your balance cannot cover the call
403	ORG_CONTEXT_MISSING	Your user has no organization to charge
429	PROVIDER_RATE_LIMITED	The model provider is busy; retry after the seconds in Retry-After
501	STREAMING_DISABLED	Streaming was requested
502, 504		The model provider failed or timed out; retry later

Authentication failures (401) and the API's own rate limit (429 without a code) answer in the rest of the API's shape, {"detail": "..."}. See Errors, pagination and limits.