AI agents that run shell commands, not just chat

What it does GoRaven is a self-hosted AI agent platform built in Go with a React frontend. It assigns each team member an isolated workspace where agents can read and write files, execute shell commands, call internal APIs and MCP tool chains, and generate reports or charts. Administrators set model quotas, tool permissions, and data access rules, while the team shares projects and a centralized skill library.

The interesting bit The project’s design philosophy is deliberately anti-hype: it treats the model as “the Way” and the agent as a deterministic instrument, focusing on observability, OS-level runtime integration, and rigid engineering discipline to prevent execution flows from derailing. That is a refreshingly mechanical take in a space crowded with “memory and evolution” buzzwords.

Key highlights

• Runs fully self-hosted via Docker; data never leaves the server and it plugs into OpenAI, Claude, DeepSeek, Qwen, GLM, or any compatible API.
• Team isolation with shared resources: individual workspaces plus shared projects, skills, and knowledge bases.
• RAG loop: agents retrieve policies, docs, and business materials in real time during execution, with cited sources.
• Skill marketplace: prompts, scripts, and workflows are packaged as reusable, one-click-installable skills.
• Plugin hooks at conversation, tool-call, and SSE lifecycle points allow deep customization without forking the core.

Verdict Teams that need autonomous, file-system-aware agents behind their own firewall should look here; solo developers wanting a drop-in chat wrapper will find the team-centric quota and workspace machinery overkill.