prismor

View on GitHub

Self-hosted runtime control plane for AI agents. Observe or HITL approve or Block rogue tool calls before it executes: secret leaks, prompt injection, supply chain etc in a local dashboard. Agent agnostic (Claude, codex, langchain etc.)

★ 38447 forksPythonApache-2.0PrismorSec

Self-hosted runtime control plane that intercepts AI agent tool calls and enforces user-written policy: observe, require human approval, or block secret leaks, prompt injection, and supply-chain risk. Agent-agnostic with hooks for Claude Code, Codex, LangChain and others plus an MCP gateway and LLM proxy.

Use Cases

Block or HITL-approve rogue agent tool calls before executionDetect and stop prompt injection in tool results and promptsPrevent secret leaks and credential exfiltration by agentsGovern MCP servers and tools through a policy-evaluating gatewayApply per-rule observe/enforce policy across agents and projectsScan skills and MCP servers for supply-chain riskGive each agent a named identity with least-privilege profilesSign and audit agent runtime telemetry locallyGuard production framework agents (LangChain, CrewAI, ADK, etc.)Proxy LLM traffic to govern agents that cannot be hooked

Built With

Language: Python
Frameworks: LangChain · LangGraph · CrewAI · OpenAI Agents SDK · Pydantic AI · AutoGen · Agno · Semantic Kernel · Google ADK · BeeAI · Claude Agent SDK · Vercel AI SDK · Mastra · browser-use · MCP · litellm

prismor

Use Cases

Built With

Tags