Introduction

Most agent frameworks help you build agents. Antraft helps you build agents you can put in production: every tool call passes a policy engine, spend is capped, content is screened for injection and data leakage, and the whole run is recorded to a tamper-evident audit trail. It works with any model from any provider and any tool source — MCP servers, REST/OpenAPI APIs, Anthropic Skills, and RAG.

Why Antraft

• Governed by default — allow / deny / pause on every action, default-deny.
• Budget as policy — enforce token and dollar spend; deny or pause when exceeded.
• Content guardrails — block prompt-injection, jailbreaks, and secret/PII leakage.
• Tamper-evident audit — a hash-chained, verifiable record of every decision.
• RBAC + multi-tenancy — scope tools by role and tenant.
• Any model, any tool — one API across providers; one adapter layer for tools.
• Cross-language — govern Python and JavaScript agents with the same engine.

Mental model

An agent proposes Actions. The runtime asks the Guard(the policy engine, exposed over a frozen JSON contract) whether each action is allowed. If it is, the tool gateway executes it — validating arguments, applying retries and timeouts, and recording the result. Guardrails screen the content flowing in and out, and the auditor writes an immutable trail.

Next steps

• Installation — install the Python package (and the TS SDK).
• Quickstart — a governed agent in a few lines.
• Architecture — how the pieces fit together.