Secure, Portable AI Agent Runtime
for Enterprise Work
Forge is the open-source runtime for Anthropic's Agent Skills standard — built for the agent that runs next to a service, in your environment, on infrastructure you already operate. Write a SKILL.md. Compile to a portable, hardened agent. Deploy it anywhere containers run: Kubernetes, on-prem, air-gapped, embedded in CI, or as an A2A endpoint.
Built on Anthropic's open Agent Skills standard · Apache 2.0 licensed · Single static binary · No managed-service dependency
brew install initializ/tap/forge Why Forge
Five things alternative runtimes have structurally chosen not to do.
Open standard, not proprietary surface
SKILL.md is Anthropic's open Agent Skills format. Forge runs that standard. AgentCore agents are AgentCore agents. Foundry agents are Foundry agents. Move runtimes and you rewrite. Move Forge skills and they just keep working.
Any infrastructure, including ones the hyperscalers can't reach
EKS, AKS, GKE, OpenShift, vanilla Kubernetes, Nomad, bare metal, on-prem, air-gapped, edge. Single binary, single container. No hyperscaler runtime is going to ship to an air-gapped facility — Forge already does.
Any model, including ones you run yourself
Anthropic, OpenAI, Gemini, Ollama, or local GGUF via forge brain. Multi-provider isn't a feature toggle — it's the default architecture. Run the model that fits the workload, including local inference for data that can't leave your network.
Production hardening as the default
Computed-not-declared trust. Per-skill egress enforcement. Denied-tools as explicit contracts. Audit logs with correlation IDs. These are part of the open-source runtime, not the enterprise tier. You don't buy the security model — you start with it.
The agent is a document, not a codebase
A SKILL.md is markdown with frontmatter — readable and authorable by SREs, security engineers, compliance officers, and any domain expert who already writes runbooks. Developers extend skills with scripts when LLM reasoning alone isn't enough. Platform engineers operate the runtime. The agent is a document, and three roles collaborate through it.
Who Forge is forForge is not trying to be every agent runtime
The agent runtime market is crowded. We respect the alternatives — and we've drawn explicit lines around what Forge is and isn't, so you can pick the right tool without reading marketing copy.
Not a desktop agent
If your agent needs to sit next to a person — completing multi-step tasks across SaaS apps, organizing files, drafting documents — use Claude Cowork. Forge is for the agent that sits next to a service, not next to a person.
Forge vs Claude Managed AgentsNot a hyperscaler-managed runtime
If your stack is 100% AWS and you want zero operational burden, use AgentCore. If you're Microsoft-native with Entra and M365, use Foundry. If you're GCP-native with Gemini grounding, use Gemini Enterprise. Forge is for organizations that don't or can't standardize on a single cloud.
Compare allNot a Python framework
If you need rich orchestration primitives, RAG composition, or LCEL graph definitions in code, use LangChain or CrewAI. Forge can run their logic inside a SKILL.md script — but Forge is a runtime, not a framework.
Forge vs LangChainNot a chat assistant
If you want a conversational AI for general questions, use Claude directly. Forge agents are headless production services, not chat interfaces. They expose Slack and Telegram channels into already-running agents — they don't replace conversation.
What is Forge
Forge compiles a SKILL.md
— Anthropic's open Agent Skills format — into a portable, hardened agent. Single static binary.
Outbound-only network posture. Computed-not-declared trust. Audit logs by default.
The same agent definition runs on your laptop, in a Docker container, as a Kubernetes deployment on EKS / AKS / GKE / on-prem, as a sidecar in a service pod, as a step in a CI pipeline, as an A2A endpoint, or in an air-gapped cluster with local model inference. No per-environment rewrite. No managed-service dependency. No vendor billing surface beyond the model API you already use.
Forge serves three audiences who collaborate through the SKILL.md document — domain experts who author skills as markdown, developers who write the tool scripts that extend them, and platform engineers who operate the runtime. The agent is a document, and the document is the contract between them.
Skills are just Markdown
Define everything in a single SKILL.md file — model, tools, egress rules, and prompt. Forge handles the rest.
---
name: code-review
description: AI-powered code review for diffs and PRs
category: developer
tags:
- code-review
- security
- pull-request
bins:
- curl
- jq
- git
env:
one_of:
- ANTHROPIC_API_KEY
- OPENAI_API_KEY
egress:
- api.anthropic.com
- api.openai.com
---
# Code Review
AI-powered code review that analyzes diffs
for bugs, security issues, and improvements.
## Constraints
- Read-only analysis, never modifies code
- Structured JSON output with severity levels
- Supports local diffs and GitHub PRsForge vs Typical Agent Frameworks
Forge focuses on infrastructure. Not just capability.
| Feature | Forge | Typical |
|---|---|---|
| Outbound-only networking | ✓ | ✗ |
| Inbound tunnel required | ✗ | Often |
| Environment portability | ✓ | Limited |
| Skill-based atomic model | ✓ | ✗ |
| Enterprise deployment focus | ✓ | Rare |
| Air-gapped compatibility | ✓ | Rare |
Designed for the Enterprise AI Workforce
AI agents are moving from assistants to workers. Workers require infrastructure, boundaries, observability, and portability. Forge provides the runtime layer for enterprise AI agents — multi-cloud, on-prem, air-gapped, regulated.
Join the community
Forge is open source and built by the community. Every contribution matters.
5 contributors and counting