What is Pikarc?

Runtime control plane for AI applications. Budget enforcement, kill switch, audit trail — at every model call and tool call.

Pikarc sits between your agent code and the models/tools it calls. Every model call and tool call passes through Pikarc, which makes an ALLOW or DENY decision based on your configured guardrails.

Core Concepts

Workspace

Your tenant boundary. Owns runs, API keys, budgets, and the kill switch. One workspace per account.

Run

A single agent execution. Contains an ordered list of Steps. Status progresses from RUNNING to COMPLETED, FAILED, or BLOCKED.

Step

One action within a run. Types include MODEL_CALL, TOOL_CALL, INPUT, OUTPUT, REASONING, and SYSTEM. Each step gets a Decision.

Decision

ALLOW or DENY, evaluated at run start and before each step. Includes the deny reason and which rules were evaluated.

How It Works

Your Agent Code
       │
       ▼
┌─────────────┐
│  Pikarc   │  ← Budget checks, kill switch, concurrency limits
│  SDK        │
└──────┬──────┘
       │ ALLOW / DENY
       ▼
  Model / Tool

1. Your agent starts a run via the SDK
2. Before each model call or tool call, Pikarc evaluates guardrails
3. If all checks pass → ALLOW — the call proceeds
4. If any check fails → DENY — PikarcBlockedError is raised
5. Token usage and costs are tracked automatically
6. The run ends, and everything is visible in the dashboard

What Gets Checked

Rule	Description
Kill switch	Emergency brake — blocks all runs instantly
User blocked	Individually blocked users
Workspace daily budget	Total spend limit per day
User daily budget	Per-user spend limit per day
Monthly run limit	Runs per month (based on plan tier)
Max concurrent runs	Simultaneous active runs

Next Steps

Quickstart

Install the SDK and wrap your first model call in 5 minutes.

SDK Reference

Full SDK API: client, guarded runs, error handling.

API Reference

REST API endpoints for runs, workspace, usage, and more.

Guardrails Guide

How the decision engine evaluates rules and denies requests.