Forge

Forge is an agent runtime built around Anthropic's Agent Skills format. You describe an agent as a SKILL.md file (markdown with YAML frontmatter), and Forge parses it, discovers the tools and dependencies it needs, compiles it into an agent specification, applies a security policy, and runs an LLM agent loop with tool calling, memory, and scheduled tasks. It ships as a single static binary and container.

The runtime targets agents that run next to a service on infrastructure you already operate: a local machine, Docker, Kubernetes distributions, on-premises and air-gapped sites, CI pipelines, or an Agent-to-Agent (A2A) endpoint with a spec-conformant agent card. Model providers include Anthropic, OpenAI, Gemini, Ollama, and local GGUF models, with automatic failover between providers.

Security controls are part of the open-source runtime: outbound-only connections with no public listeners, per-skill egress allowlists enforced at build time and runtime, encrypted secrets, denied-tool contracts, Ed25519 artifact signing, content guardrails with PII detection, and structured audit logs with correlation IDs. Authentication plugs into OIDC, AWS SigV4, GCP IAP, and Azure AD, and OpenTelemetry tracing links spans to audit events.

Agents reach users through Slack and Telegram connectors, connect to remote HTTP MCP servers such as Linear, Notion, and Atlassian, and can be managed in a web dashboard with a skill builder. An interactive builder can also open Claude Code with Forge's MCP tools. The core, skills, and plugin modules can be imported as Go libraries.