prismor

Prismor is a self-hosted runtime security and control plane for AI agents. It intercepts tool calls before execution and applies user-defined policies to observe, require human approval, or block actions such as secret leaks, prompt injection, destructive commands, privilege escalation, and supply-chain risks. The project is agent-agnostic and documents integrations with coding agents including Claude Code, Codex, Gemini CLI, Cursor, GitHub Copilot, OpenCode, and many others, as well as production frameworks such as OpenAI Agents SDK, LangChain/LangGraph, CrewAI, Pydantic AI, AutoGen Core, and Semantic Kernel. The system uses a single policy engine behind multiple enforcement surfaces: runtime hooks, an MCP gateway, framework adapters, an LLM proxy, and Claude inference hooks. Policies are written in YAML with per-rule observe or enforce modes; the default is observe, so nothing blocks until rules are explicitly enabled. Prismor also provides supply-chain checks, network isolation, MCP guardrails, prompt guardrails, tool tags, secret cloaking, canary honeytokens, IAM identities, semantic guard, session learning, and a signed audit trail. A local web and terminal dashboard shows agent activity, session forensics, token usage, and findings. An optional self-hosted control plane supports signed remote policy, device enrollment, and redacted telemetry. Installation is via pip with `prismor setup`, and the project includes extensive documentation for governance modes, CLI commands, Docker use, and compliance-oriented attestation bundles.