OpShield
An open-source security middleware platform for AI agents. By intercepting tool calls, it provides risk classification, budget control, state rollback and chaos testing capabilities. It can be self-hosted and has zero dependencies.
Timeline (last 90 days)
OpShield is an open-source security platform that serves as security middleware for AI agents, intercepting every tool call and providing capabilities such as risk classification, budget control, state rollback, and chaos testing.
The background of OpShield's creation is that after Guardrails AI was acquired by Harvey, the market lacked an independent, self-hostable security alternative.
OpShield works at the tool-call layer, where each tool call passes through a risk classification engine, is assigned a level from LOW to CRITICAL, and high-risk calls can be automatically blocked.
OpShield provides budget enforcement and cost prediction; before the agent calls a tool, the system estimates the cost and compares it with budget thresholds.
OpShield has a built-in circuit breaker that automatically cuts off calls to a tool when the tool fails repeatedly, preventing the agent from falling into ineffective retries.
OpShield supports "one-call rollback" through state snapshots, allowing developers to quickly restore to a previous known state.
OpShield has built-in OpenTelemetry-compatible distributed tracing, as well as an agent scoring system (0-100 points, A-F grades), and includes a chaos engineering module.
OpShield claims to be "zero-dependency" and can be used with OpenAI, Anthropic, LangChain, CrewAI, AutoGen, or any Python function.
OpShield has been open-sourced on GitHub and can be installed via pip install opshield.
Current information about OpShield mainly comes from the author's own release post, and no third-party independent evaluations or production environment cases have been seen yet.
All Facts (10)
Current information about OpShield mainly comes from the author's own release post, and no third-party independent evaluations or production environment cases have been seen yet.
50%UnverifiedThe background of OpShield's creation is that after Guardrails AI was acquired by Harvey, the market lacked an independent, self-hostable security alternative.
50%UnverifiedOpShield works at the tool-call layer, where each tool call passes through a risk classification engine, is assigned a level from LOW to CRITICAL, and high-risk calls can be automatically blocked.
50%UnverifiedOpShield provides budget enforcement and cost prediction; before the agent calls a tool, the system estimates the cost and compares it with budget thresholds.
50%UnverifiedOpShield has a built-in circuit breaker that automatically cuts off calls to a tool when the tool fails repeatedly, preventing the agent from falling into ineffective retries.
50%UnverifiedOpShield supports "one-call rollback" through state snapshots, allowing developers to quickly restore to a previous known state.
50%UnverifiedOpShield has built-in OpenTelemetry-compatible distributed tracing, as well as an agent scoring system (0-100 points, A-F grades), and includes a chaos engineering module.
50%UnverifiedOpShield claims to be "zero-dependency" and can be used with OpenAI, Anthropic, LangChain, CrewAI, AutoGen, or any Python function.
50%UnverifiedOpShield is an open-source security platform that serves as security middleware for AI agents, intercepting every tool call and providing capabilities such as risk classification, budget control, state rollback, and chaos testing.
50%UnverifiedOpShield has been open-sourced on GitHub and can be installed via pip install opshield.
50%