securevector-guard

• Secures tool calls and LLM traffic — locally. Allow / deny / ask, enforced at agent runtime via native plugins, framework SDKs, or the multi-provider proxy. Nothing leaves your machine.
• Tamper-evident audit chain — every tool call appended to a SHA-256 hash-chained log, verifiable from the Tool Activity tab.
• 72 detection rules + Guardian ML — OWASP LLM Top 10 + 28 agent-attack chains (prompt injection, jailbreaks, credential exfiltration, PII). An offline ML model catches obfuscated/encoded attacks regex misses — sub-millisecond, fail-open. Details ↓
• Token + cost tracking — per-agent, per-model spend in real time.
• Apache 2.0, no signup — pip install and you're covered in 60 seconds.

Works with every agent

Agent / runtime	How to add	Audit runtime_kind
Claude Code	Native plugin — inline hooks, zero proxy	claude-code
OpenAI Codex (CLI 0.133+)	Native plugin	codex
GitHub Copilot CLI	Native plugin	copilot-cli
Cursor	Native plugin	cursor
OpenClaw / ClawdBot	Native plugin	openclaw
LangChain	SDK — securevector-sdk-langchain (pip install)	langchain
LangGraph	SDK — securevector-sdk-langgraph (pip install)	langgraph
CrewAI	SDK — securevector-sdk-crewai (pip install)	crewai

Native plugins enforce inline (zero proxy); SDKs secure tool calls — all on one enforcement core, so a single tool_id="Bash" rule covers Bash on Claude Code, exec_command on Codex, and shell calls on Cursor & OpenClaw. SDK modes: observe logs (tool runs), enforce blocks. Install from the Integrations tab.

Also supported for LLM traffic (no per-agent runtime_kind attribution): the multi-provider proxy — Ollama · n8n · any OpenAI-compatible app via OPENAI_BASE_URL; the MCP server — Claude Desktop & other MCP clients (pip install securevector-ai-monitor[mcp]); and a raw POST /analyze for any HTTP client. MCP tools invoked from a plugged-in harness are attributed to that harness's runtime_kind.

What's new in v4.8.0

• Framework SDKs for LangChain, LangGraph, and CrewAI — pip-installable middleware/tool wrappers that secure tool calls and stream them onto the Agent Map by runtime_kind (see the "Works with every agent" table above).

What's new in v4.7.0

• Fleet management (cloud accounts, optional) — fleet-wide Agent Maps + Agent Runs for devices enrolled via a mint token (SVET); opt-in and metadata-only. Non-enrolled (local-only) installs forward nothing.
• SecureVector Guard for Cursor — native plugin + hooks for the Cursor agent (see the plugins table above).

Full release history in the CHANGELOG.

How It Works