NEWv2.4.2 — reference impl. of the MI9 protocol

The governed framework for agents.

Antraft gives your agents state-of-the-art governance, policy enforcement, budget controls, and audit trails — all built in. Works with any model.

Start Building

Govern every model stack

AAnthropicOAOpenAIGGeminiBBedrockAzAzureVxVertexCoCohereMMistralGqGroqOlOllama

16µs

governance overhead / eval

2.4×

parallel tool speedup

100+

models, every provider

secret leaks (CI-enforced)

What you get

Governance is the substrate, not a plugin.

LangChain and LangGraph give you agents. Antraft gives you agents you can put in production — every action accountable.

Governed by default

Every tool call passes the policy engine — allow, deny, or pause. Default-deny, so agents can only do what you permit.

Budget as policy

Enforce token and dollar spend as a rule. Deny on a hard cap, or pause for human approval on a soft one.

Content guardrails

Block prompt-injection in tool output, jailbreaks in input, and secret or PII leakage in either direction.

Tamper-evident audit

A hash-chained, append-only audit trail. Any edit, reorder, or deletion is cryptographically detectable.

Any model, any company

Anthropic, OpenAI, Gemini, Bedrock, Azure, Vertex, Cohere, Mistral, Groq, Ollama — one API, all governed.

Any tool

MCP servers (local and remote), REST / OpenAPI, Anthropic Skills, and RAG — every source is policy-controlled.

Collaboration in Action

Built for teams and modern creators.

Whether orchestrating complex multi-agent graphs or launching a fast-growing solopreneur project, Antraft provides the workspace and governance engines to scale securely.

Workflow Orchestration

Teaming up agents with zero friction

Define secure channels, permission models, and budget limits for multi-agent conversations. When agents work together, Antraft guarantees every interaction is verified, audited, and strictly controlled.

Solopreneurs

Launch fast, stay compliant

Build and test secure agent workflows in minutes. Get enterprise-grade safety out of the box so you can focus on building value, not compliance plumbing.

Team Alignment

Cohesive team workspace

Coordinate agent designs, control API keys, and monitor collective execution logs in a secure, central interface tailored for team collaboration.

Creative team collaborating with AI agents

Creative Agency

Unlocking collective creativity

Bring designers, developers, and AI agents into the same workflow. Antraft provides clean abstractions to build interactive workspaces where humans and agents co-create.

Goal alignment and strategy illustration

Shared Objectives

Aligning goals & tasks

Map out high-level strategies and let your agent swarm execute them safely under strict guardrails. Watch agents coordinate to solve complex problems with complete transparency.

The difference

Built for production, not demos.

CapabilityAntraftOthers

Build LLM agents & multi-agent graphs✓✓

Policy-enforced on every tool call✓—

Budget (token + $) enforced as policy✓—

Content guardrails (injection / PII / secrets)✓—

Tamper-evident audit log✓—

RBAC + multi-tenancy✓—

Cross-language governance (Python + JS)✓—

Quickstart

Governed in a few lines.

Pick any model, register any tools, set the rules. The runtime enforces them on every step — and the same engine governs your JavaScript agents too.

→Default-deny policy — opt tools in explicitly
→Budget, retries, timeouts, checkpoints built in
→Audit, replay, and a live observability dashboard

agent.py

from antraft import Antraft, ToolRegistry, tool_from_function
from antraft.models import create_model
 
def search_web(query: str) -> str:
    "Search the web."
    return f"results for {query}"
 
tools = ToolRegistry([tool_from_function(search_web)])
 
runtime = (
    Antraft.agent(
        create_model("anthropic:claude-opus-4-8"),
        tools,
        task="Research Antraft and summarize it.",
        guardrails=True,
    )
    .allow(["search_web"])       # default-deny everything else
    .budget(max_cost_usd=1.0)    # enforce spend as policy
    .build()
)
 
await runtime.run()              # every tool call is governed + audited

Ship agents you can trust.

Open source. MIT licensed. Governed by default.

$ pip install antraftStar on GitHub