What is ForgeOS?

ForgeOS is an open-source Skill Intelligence operating system for AI agents, helping agents choose the right technique, use minimal context, execute in a controlled manner, and only acknowledge task completion when there is verification evidence.

Instead of merely providing additional prompts or tools, ForgeOS acts as a control plane surrounding the agent, skills, context, policy, execution graph, and evidence.

Problems that ForgeOS Solves

Current AI agents can plan and use various tools, but often face issues such as:

• Choosing the wrong skill or tool.
• Injecting too much unnecessary context.
• Difficulty explaining why an action was taken.
• Lack of evidence to prove the accuracy of results.
• Challenges in recovering or continuing after a workflow error.
• Risk of learning from incorrect data and inadvertently degrading system quality.

ForgeOS is designed to transform the agent's execution process into a clear, verifiable, and recoverable stream.

Key Features

Unified Skill Intelligence Router

ForgeOS analyzes goals, constraints, tools, reliability, and context to select the appropriate technique. RoutePlan records both the reasons for selection and the reasons for discarding unsuitable options.

Global Context Kernel

Automatically generates a ContextPack for each work unit, loading only the data that is truly necessary. Discarded sources are logged for audit purposes.

Deterministic Skill Fabric

Workflows are compiled into an execution graph combining various types of steps:

• Deterministic: rules, scope, coverage, and evidence.
• Agent: investigation, analysis, and assessment.
• Reflection: detecting contradictions and reducing false positives.
• Control: retry, gate, rollback, and recovery.

Evidence-gated delivery

Each step can be associated with evidence, evaluation criteria, and verification status. Outcomes are only released once they pass the appropriate verification gate.

Continuous Learning with quarantine

New knowledge is not automatically injected directly into the system. New patterns are isolated, tracked from their origin, limited in their lifespan, and require an independent evaluator before being upgraded.

Agent Surface Security

ForgeOS can review instructions, skills, hooks, MCP configurations, access rights, package lifecycle, environmental variables, and paths that risk data leakage.

Recovery and audit

The system supports receipt, provenance, lease, fencing, snapshots, rollback, and audit trails so that workflows can continue or recover safely after errors.

Integration Capabilities

ForgeOS provides adapters and profiles for various agent environments, including:

• ChatGPT
• Codex
• Claude Code
• Cursor
• Cline
• Continue
• Gemini CLI
• OpenClaw
• OpenCode
• Hermes
• Custom agents

The system supports integration protocols and surfaces such as MCP, A2A, CLI, HTTP, and Studio Dashboard.

ForgeOS v0.6.1 is now available

• 128 deep kernel techniques
• 32 L0 techniques for coordination, trust, and context
• 96 L1 cross-domain techniques
• 1,024 typed outcome scaffolds
• 128 independent evaluator bindings
• Router Precision@1 achieving 93.75% on current benchmarks
• Router Precision@3 and Recall@6 achieving 100%
• Unsafe route activation achieving 0% in the current test suite
• Agent-surface adversarial test achieving 20/20

Note: 1,024 outcome scaffolds do not equate to 1,024 production-grade skills. ForgeOS publicly documents the boundaries of verification and does not claim capabilities that lack sufficient evidence.

Quick