Open source

Build agents in plain English.

Forge is the open-source, secure, portable AI agent runtime for enterprise work — the runtime for Anthropic’s Agent Skills standard. Write a SKILL.md, compile it to a portable, hardened agent, and deploy it anywhere containers run. Free, Apache-2.0, distributed via useforge.ai — and governed at enterprise scale by initializ.

Get Forge at useforge.ai → Star Forge on GitHub

16 embedded skills23+ built-in tools5 LLM providers + customApache 2.0A2A 0.3.0

Running in sixty seconds One static binary — no managed-service dependency.

Install, init, run.

Forge ships as a single static binary. forge init scaffolds an agent — forge.yaml, SKILL.md, .env — and forge run starts it locally, in Docker, on Kubernetes, or air-gapped. Add --with slack and it’s in your team’s channel.

terminal

brew install initializ/tap/forge
forge init my-agent
cd my-agent
forge run --with slack

The agent is a document Describe what your agent should do; Forge compiles it into a hardened agent.

SKILL.md — versioned, reviewable, portable.

A skill is a markdown file: YAML frontmatter declares what the agent needs — binaries, egress domains, denied tools — and plain-English instructions follow. It goes through code review like everything else you ship, and it runs the same everywhere. Sixteen embedded skills come in the box, from code review to Kubernetes incident triage.

SKILL.md

---
name: k8s-incident-triage
category: sre
description: Triage Kubernetes incidents
metadata:
  forge:
    requires:
      bins: [kubectl, jq]
    egress_domains:
      - api.pagerduty.com
    denied_tools: [file_write]
---
# Incident Triage
You are an SRE triage agent...

Secure by default Plus a three-layer platform policy — system, user, workspace — where the most restrictive setting wins.

Outbound-only by design

Channels connect outward — no public webhooks, no tunnels, zero inbound attack surface.

Deny-all egress

Domain allowlists enforced at build time and at runtime. Unlisted destinations are blocked by default.

Guardrails engine

PII detection, jailbreak and prompt-injection detection, command-injection checks, secret detection, and custom regex/keyword rules.

Hardened binary execution

13 layers of control on every CLI call — shell denylist, binary allowlist, argument validation, path confinement, environment isolation.

DEFER: approvals in chat

High-risk tool calls pause for a human decision delivered to your channel; unanswered approvals auto-deny.

Audit you can verify

A hash-chained, append-only event log — run forge audit verify and any tampering is detectable.

Managed identity for MCP Agents don’t hold long-lived MCP credentials — tokens are brokered, scoped, and expiring.

Platform identity or user identity — per server.

Every MCP server gets its own token resolver. type: platform gives the agent its own scoped, short-lived token — the refresh credential stays platform-side. type: user acts as the requesting person: each user consents once and gets their own token, so writes are attributed to the human who asked.

forge.yaml

mcp:
  servers:
    - name: atlassian-read
      url: https://mcp.atlassian.com/mcp
      auth: { type: platform, ref: mcp.atlassian }
      # agent-scoped, short-lived token
    - name: atlassian-write
      url: https://mcp.atlassian.com/mcp
      auth: { type: user, ref: mcp.atlassian }
      # per-user delegated access, on consent

Meet your stack where it is

Channels

Connect agents to Slack and Microsoft Teams — outbound-only connections, no exposed endpoints.

Models

OpenAI, Anthropic, Gemini, Ollama, or any custom endpoint — including AWS Bedrock (SigV4) and enterprise gateways like Kong — with automatic fallback chains.

Build open. Govern at scale.

From Forge to production.

initializ is the enterprise control plane for open-source Forge AI agents — built with AgentSkills in plain English — giving organizations the governance, security, and observability to run them at scale. When your Forge agents need policy, RBAC, and tamper-evident audit, that’s initializ.

Govern Forge agents with initializ →

Forge questions

What is Forge?

Forge is the open-source framework for building AI agents with AgentSkills written in plain English. It’s free, community-driven, and distributed via useforge.ai.

What are AgentSkills?

AgentSkills are agent capabilities defined in plain English — readable, composable descriptions of what an agent can do, which Forge turns into working skills.

Is Forge free?

Yes. Forge is open source and community-driven. You can build and run agents with it free, via useforge.ai and GitHub.

How does Forge relate to initializ?

Forge is the open-source framework for building agents. initializ is the commercial control plane that governs Forge agents — and other frameworks — at enterprise scale, with policy, security, observability, and audit.

Do Forge agents require initializ?

No. Forge stands on its own for building and running agents. initializ adds enterprise governance — policy, RBAC, observability, and tamper-evident audit — when you run agents at scale.

Which channels does Forge support?

Forge connects agents to Slack and Microsoft Teams through outbound-only connections — no public webhooks, no tunnels, zero inbound attack surface.

Build fast. Own your agents.

Start with Forge today — and bring initializ when you’re ready for scale.

Get Forge → Book a Demo