FireGuard

FireGuard Concepts

Understand conversations, guardrail checks, execution modes, and monitoring.

FireGuard sits between your application and your model. It checks input before the model runs and output before the response reaches the user.

Request Flow

Conversations

A conversation groups input and output checks into one thread. Create one conversation per user-visible chat, ticket, workflow run, or agent session.

Project Configuration

The API follows the guardrails configured on your FireGuard project:

  • Policies can block on input, output, or both.
  • Security guardrails can check input, output, or both.
  • Topics power monitoring and analytics.

Your request can optionally include guardrails, but the project configuration determines what actually runs.

See Project configuration for policies, security guardrails, topics, project IDs, and the sandbox workflow.

Agent Hooks

FireGuard Agent Hooks adapt the same input and output guardrail flow to AI coding agents. They create or reuse a FireGuard conversation for each agent session, run input checks before prompts, commands, tool calls, and file access, then audit supported outputs.

Execution Modes

ModeUse whenBehavior
normalYou want complete response-path resultsRuns guardrails sequentially and returns full policy/security details.
fastYou want lower latencyRuns eligible checks in parallel and returns as soon as a blocking result is known.

Set execution_mode in the request body to override the project default for one call.

Monitoring

FireGuard stores analyzed turns for the monitoring dashboard and export API. Use the monitoring endpoint to pull message rows, policy metrics, and security issue counts into your own dashboards.

Usage & Billing

Long text, images, and organization limits can affect request units. See Usage & billing for the unit formula, image behavior, daily limits, and credit errors.

On this page