OpenViking 🔌 MCP server Open source

What it does

OpenViking, from Volcengine, is an open-source context database for AI agents. It gives an agent one place to keep project knowledge, user memories and reusable skills across sessions. Everything is organised as a virtual filesystem under viking:// URIs, so an agent browses, reads, writes and searches context the same way it works with files.

What is inside

Content is split into resources such as docs and repositories, per-user memories, and skills. Each processed directory gets a one-line abstract and a short overview, so an agent can judge relevance before loading the full content. Retrieval combines vector search to find candidate directories with exploration inside them. Committing a session archives the conversation and extracts memories in the background, with policies that decide what to keep, merge or skip. The package includes a server, an ov CLI, Python, Go and TypeScript SDKs, an HTTP API, a web studio, a beta desktop console, and an optional agent framework called VikingBot. It needs an embedding model and a vision-language model, from a cloud provider or local Ollama.

Works with

The docs list native integrations using hooks plus MCP for Claude Code, Codex, Cursor and TRAE, a plugin for OpenCode, and extensions for Pi, OpenClaw, Hermes, DeerFlow and LangChain. Any MCP client can use its memory and context tools.

How to install

pip install openviking --upgrade
openviking-server init
openviking-server doctor
openviking-server

Maintenance and safety

The main server is AGPL-3.0, with the CLI and examples under Apache-2.0, and it is updated daily by a large team. A commercial edition exists, but the self-hosted server needs no activation key. It stores conversation content and memories, so configure authentication before exposing it beyond localhost and check the AGPL terms if you plan to offer it as a service. The published benchmarks are the project's own.

Who should use it

Developers who want their coding agents to remember preferences, project context and past fixes across sessions, or who are building their own agents and need a structured memory and retrieval layer. It is more infrastructure than a quick plugin.