This is an early release preview. You may encounter bugs.
Forge logo
Unclaimed

Tool deployment security

Forge

Forge is the open-source runtime for Anthropic's Agent Skills standard — built for the agent that runs next to a service, in your environmen

B 77/100 GitHub score ? This grade is derived from GitHub signals, not user votes. Open for the full breakdown.
No votes yet

01 / About

What Forge is.

Forge is a runtime for agents defined as Agent Skills: you write a SKILL.md file with YAML frontmatter, and Forge parses it, discovers the tools and requirements it declares, compiles an agent specification, applies a security policy, and runs the LLM agent loop with tool calling, memory, and scheduled tasks.

The runtime ships as a single static binary in Go and is designed to run wherever containers run — locally, in Docker, on Kubernetes, on-premises, in air-gapped facilities, embedded in CI, or as an agent-to-agent (A2A) endpoint that serves a spec-conformant Agent Card at /.well-known/agent-card.json. Model providers are pluggable: Anthropic, OpenAI, Gemini, Ollama, and local GGUF weights are all supported, with fallback chains between them.

Security controls are part of the runtime rather than an add-on tier: outbound-only connections, build-time and runtime egress allowlists enforced through a subprocess proxy, encrypted secrets, Ed25519 signing and verification of build artifacts, pluggable authentication providers (OIDC, AWS SigV4, GCP IAP, Azure AD), content filtering and PII guardrails, and structured NDJSON audit logs carrying correlation IDs for every action.

Features

  • Skill-defined agents: SKILL.md files with YAML frontmatter compile into runnable agents, extensible with scripts when reasoning alone is not enough
  • Egress security: domain allowlists enforced at build time and runtime, with a subprocess proxy and denied-tools contracts
  • Channel connectors: Slack via Socket Mode and Telegram via polling, both outbound-only
  • Cron scheduling: recurring tasks with delivery to a channel, and a hybrid file/CronJob scheduler backend on Kubernetes
  • Memory: session persistence plus long-term vector search, with reversible compression of bulky tool outputs
  • Model providers: Anthropic, OpenAI, Gemini, Ollama, and local GGUF weights, with automatic failover between providers
  • MCP client: connects to any HTTP Model Context Protocol server, surfacing tools as with namespaced audit entries
  • Observability: OpenTelemetry tracing with span propagation, cross-linked to the audit log
  • Web dashboard: forge ui provides browser-based agent management and a skill builder
  • Library modules: forge-core, forge-skills, and forge-plugins can be imported as Go libraries

02 / Discussion CREDIBILITY-GATED

Discussion

Reading is open to everyone. Posting and voting need a verified identity or a GitHub grade of B or higher.

  • No discussions yet.

03 / Build

Build with Forge.

Browse the catalogue for frameworks, tools, and harnesses, each scored on real GitHub credibility.

Get Forge →

Browse the catalogue