Open Source · Self-Hosted · AI Harness
One runtime. Models, tools, knowledge, skills, and workflows — assembled so your Agent delivers results, not just replies.
Website · Live Preview · Discord · 中文
AI chat is everywhere. What's hard is turning it into an engineering system your team can actually rely on.
Raven is an AI Harness — every team member gets their own workspace, and inside it the Agent reads files, writes code, runs commands, calls APIs, and searches your knowledge base. It takes on tasks, breaks them down, picks the right tools, and delivers results — instead of spitting out text for you to copy-paste.
You describe the goal → Agent understands it → Decomposes into subtasks → Runs in parallel → Delivers results
↓
Reads docs · Writes code · Calls APIs · Searches knowledge · Installs deps
From "model outputs text" to "agent delivers outcomes." That's what Raven does.
|
Every member gets an isolated workspace. The team shares projects and skill libraries. Admins control model quotas, tool permissions, and data access — all in one place. |
Agents don't stop at suggestions — they read and write files, run shell commands, call MCP tools, connecting your internal APIs, databases, and private services directly. |
|
Prompts, scripts, and workflows packaged as installable skills. One-click install, automatic dependency resolution, centralized versioning. One person figures it out — the whole team skips the hard part. |
Policies, docs, and business data fed into RAG. Agents retrieve context in real time. Knowledge stops being static documentation and becomes the basis for every decision. |
|
Before/after conversations, during tool calls, on event streams — lifecycle hooks let you inject custom logic anywhere. Deep customization without touching a line of core code. |
One Docker command and you're running. Model data never leaves your infrastructure. Works with OpenAI, Claude, DeepSeek, Qwen, GLM, and compatible APIs. |
docker pull 8treenet/raven:latest
docker run -d --restart=always --name raven-agent \
-p 8000:8000 \
8treenet/raven:latestdocker run -d --restart=always --name raven-agent \
-p 8000:8000 \
-v /opt/raven_data:/raven/data \
8treenet/raven:latestAfter startup, visit http://localhost:8000 and follow the setup wizard to create your admin account.
The container uses UTC by default. To set a different timezone, add
-e TZ=America/New_York. Available timezones: https://en.wikipedia.org/wiki/List_of_tz_database_time_zones
| Capability | Description |
|---|---|
| 🧠 Model Orchestration | No vendor lock-in. Tasks get routed to the right model — expensive ones for precision work, cheap ones for the heavy lifting. Every token counts. |
| ⚡ Agent Orchestration | Complex tasks auto-decomposed, multiple agents running in parallel, context converging on its own. Like an AI project manager running a micro engineering team — you set the goal, it handles the rest. |
| 🎯 Skill Marketplace | Battle-tested prompts, handy scripts, proven workflows — packaged as skills, one-click install for the whole team. One person's experience becomes the team's muscle memory. |
| 👥 Team Collaboration | Everyone gets their own workspace, agents operate in sandboxed environments. Team spaces share skills, files, and sessions — admins keep control, members stay in their lane, no one steps on each other. |
| 📖 Knowledge Base | Feed in docs, specs, and business know-how. When your agent answers, it pulls real sources — no more hallucinating from memory. Every claim has a receipt. |
| 🛡️ Security Sandbox | Fully isolated command execution — minimal permissions, network controls, independent filesystems. Even third-party skills can't touch your host. Security is the foundation, not afterthought. |
| 🔌 MCP Toolchain | Internal APIs, databases, private services, CLI tools — connect them all. Your agent doesn't just suggest — it queries data, invokes services, and takes action. |
| 🪝 Plugin Hooks | Inject logic at any lifecycle point — before/after conversations, on tool calls, on event streams. Deep customization without touching a line of core code. |
| 📊 Operations Dashboard | Who's using what, which model costs the most, how active is your team — one panel to see it all. AI spending you can actually account for. |
Raven is not a KPI project, and it never will be.
Every feature, every line of code, even a single button — it earns its place, or it doesn't ship. We build a sharp, reliable tool, not a bloated platform padded with checkboxes to please a roadmap committee.
No feature creep. No enterprise vanity.
We don't chase buzzwords or ride hype cycles. What matters to us is engineering quality and whether the product actually feels good to use. You open Raven, and where things are and how they work should be obvious at a glance — that's our definition of "good."
Just a team agent platform that does one thing well.
- Core code open-sourced
- Team collaboration
- Security sandbox
- Open API
- 💬 Discord — feedback, discussions, tips
- 🌐 8tree.net — Website
- 🚀 preview.8tree.net — Live preview