GoRavenopen source · self-hosted · ai harness
GoRaven
Not a chat window. A workstation.
GoRaven gives every teammate an independent agent workspace. Reads files, writes code, runs commands, calls your APIs, searches your knowledge base — you assign the task, it plans the steps, works in parallel, delivers the result.
quick start
Visit http://localhost:8000 to start the initialization process. Create an admin account to get started.
Most AI tools hand you a paragraph to copy-paste. GoRaven hands you the outcome. The agent takes the task, plans the steps, spins up sub-agents, writes real files, runs real commands. What comes back is a report on your disk, a chart in your browser, a deployed service — not another suggestion you have to act on yourself.
design principles
The model is the Way; the agent is the instrument. The Way evolves on its own — the instrument need only be precise.
We don’t do “memory and evolution” — if it can’t be distilled into generalizable rules or real-time retrieval, it’s just noise. The model’s evolution is the model’s own business; the agent shouldn’t overstep. Its duties are three: deterministic workflows, precise tool calls, robust execution. Borrow the instrument to let wisdom arise — a true instrument fulfills its purpose to the fullest.
Observability First
Collaboration isn’t about delegating tasks — it’s about aligning understanding. Human-agent communication is far more fragile than human-human: you can’t see what it’s thinking, and it can’t guess what you want. Observability of the execution process must be pursued at the product level — every round of observation becomes the reserve for the next round of decisions.
OS First
Models have knowledge and reasoning, but no eyes and no hands. For an agent to get work done, it must lean heavily on the operating system. But an agent can’t ship with a full OS — so runtime and dependency integration must come first. This isn’t a code problem or an algorithm problem. A good agent is, at its core, an Agent OS — except you can’t actually build an operating system from scratch.
Engineering First
Agents are too flexible. Tool combinations explode, and tweaking one pruning rule sends the whole flow off track. Without solid engineering, a few iterations in and nobody dares touch it. The time spent on engineering isn’t about shipping one more feature today — it’s about making sure the project is still worth working on next year.
why goraven
Every agent can chat. Few can work.
Team-first, not single-player
Everyone gets an isolated workspace. The team shares projects and skill libraries. Model quotas, tool permissions, data boundaries — admins control it all. This is infrastructure for a whole team.
Actually does things
Agents read and write files, execute shell, call MCP tool chains. Internal APIs, databases, private services — all reachable. You get actions taken, not suggestions written.
Skills turn experience into assets
Prompts, scripts, workflows get packaged as skills. One-click install, centralized maintenance. One person hits a wall — the whole team skips it next time.
Knowledge enters the execution
Policies, docs, API references, business materials go into LLMWiki. The agent retrieves context while planning, coding, debugging. Answers have sources, not hallucinations.
Sub-agents carry the long work
The main agent understands the goal and delegates well-scoped subtasks in parallel. Complex work no longer dies in one conversation — it becomes a process you can advance, summarize, and review.
Hooks reach every stage
Before and after conversations, around tool calls, inside SSE event streams — inject custom logic at any lifecycle point. Deep customization without forking core code.
One command, your server
Data never leaves your machine. Works with OpenAI, Claude, DeepSeek, Qwen, GLM, or any compatible API. Models are schedulable resources — pick by task, balance cost and quality.
Operations stay visible
Token trends, model distribution, user rankings, tool usage, abnormal-spend alerts — one dashboard. See how the team actually uses AI, then tune the whole system.
architecture
Five layers, one runtime. Access, orchestration, capability, model, storage — each replaceable, all assembled. The value is not in any single layer; it is in the chain.
The point was never the agent. It’s how much real work it can reach.
GoRaven doesn’t invent agents. It puts them into your server, your workflow, your team’s daily routine. Open-source, self-hosted, data stays home. Simple enough to run with one command; complete enough to run for years.
View Source