Repository navigation
Security
LightAgent is intentionally flexible: it can call Python tools, execute built-in utilities, connect MCP servers, retrieve memory, and activate Skills. Production deployments should treat these extension points as trust boundaries.
Tools can read files, call networks, write databases, upload objects, or execute code depending on the callable implementation.
Recommended controls:
- Keep tool descriptions narrow and accurate.
- Validate required arguments and types.
- Avoid exposing high-risk tools to broad user traffic.
- Add approval steps for file writes, payments, database mutations, or external actions.
- Log tool name, arguments metadata, user id, and result status.
- Do not log secrets or sensitive raw payloads unless your compliance process requires it.
Legacy arbitrary-Python utilities are not registered by default in v1.0.0 and are blocked in model/tool dispatch. A direct trusted call is not a sandbox and is not safe for untrusted input. Use safe_expression for bounded arithmetic, or an independently isolated backend when controlled execution is required. See the v1.0.0 safety guide for authorization, fail-closed hooks, approvals, trace redaction, and optional signed Providers.
Memory is a persistence boundary. Use MemoryPolicy, namespaces, provenance metadata, and backend-level access control.
Recommended defaults:
- separate namespaces per user, tenant, agent, and environment
- filter retrieval by provenance and user id
- treat retrieved memory as untrusted context
- do not allow one user's memory to overwrite trusted facts for another user
See Memory and the tagged Memory Security Guidance.
MCP servers can expose external systems. For each server:
- verify command paths and server code
- keep credentials out of committed config files
- use least-privilege tokens
- disable unused servers
- test the server independently before connecting it to an agent
Skills can load long instructions, read references, execute scripts, and access assets.
Recommended controls:
- review third-party Skills before use
- keep scripts in
scripts/small and deterministic - avoid broad filesystem assumptions
- do not put secrets in
SKILL.md - prefer references for static knowledge and tools for external actions
Structured trace events are designed to avoid storing full prompts by default, but logs and tool outputs can still contain sensitive content.
Recommended controls:
- keep trace exports out of public issue reports unless reviewed
- redact secrets from debug logs
- avoid logging full memory contents
- rotate logs in long-running deployments
Report sensitive details through GitHub Private Vulnerability Reporting. See the repository's security policy.
LightAgent Wiki - see the repository, releases, and issues.
- Installation and Your First Agent
- Tools and Integrations
- Memory, Sessions, and RAG
- Workflows and Multiple Agents
- Agent Backends and Codex
- Durable Execution, Jobs, and Subagents
- Safety, Observability, and Evaluation
- Usage, Cost Estimates, and Budgets
- Deployment, Migration, and Verification
- 安装与第一个 Agent
- 工具和集成
- 记忆、会话与 RAG
- 工作流与多 Agent
- 执行后端与 Codex
- 持久执行、Jobs 与 Subagents
- 安全、观测与评测
- 用量统计、费用估算与预算
- 部署、升级与验收
- Quick Start
- Core Concepts
- API Reference
- Examples Cookbook
- Migration Guide
- Tools
- Tool Generator
- Memory
- MCP
- Skills
- LightFlow
- Tree of Thought
- Self-Learning
- Multi-Agent
- Tracing and Debugging
- Langfuse Observability
- Model Providers
- browser-use Integration
- Testing and CI
- Deployment Guide
- Architecture
- Security
- Known Limitations
- FAQ
- FAQ 中文
- Roadmap
- Release Process
- Contributing