AgenticOS Offers a Self-Hosted Open-Source Platform for Enterprise AI Agents

Summary

AgenticOS is an Apache-2.0, self-hosted platform from Vstorm for building, running, and governing a company’s AI agents on PostgreSQL and Docker. It combines agent specifications, model profiles, document retrieval, context files, reusable skills, browser access, Python sandboxes, charts, image generation, delegation, and MCP tool connections. The project says its catalogue mirrors 5,802 MCP tool servers, with 99 manually checked and their OAuth flows wired, while agents can also connect to any MCP server by URL. A quick-start script brings up the service with Docker and supports OpenAI, Anthropic, Google, and OpenRouter or postpones provider selection; published images target amd64 and arm64. Governance features include per-agent budgets checked before each model request, step limits, approval gates for side-effecting actions, encrypted organization-separated secrets, tenant isolation, versioned publishing, and an audit trail recording questions, tool calls, duration, cost, and approvals. The same runner can expose an agent through web chat, a hosted page, widget, HTTP API, WebSocket, Slack, Telegram, Mattermost, schedules, and triggers. The stack uses FastAPI, PostgreSQL with pgvector, Redis, Prefect, Pydantic AI, and Next.js 15, and the repository states that outbound calls are limited to those made by agents. The comparison section identifies evaluation harnesses, SAML/SCIM, and some connector capabilities as roadmap items, so the platform’s broader “operating system” claim includes stated limitations.