OpenViking
Context database for AI agents that stores memory, knowledge and skills as files
About OpenViking
OpenViking is a context database for AI agents. The problem it addresses is that an agent's context is normally scattered: some of it in a vector store, some in markdown files nobody maintains, some in the chat history that vanishes when the session ends. OpenViking gathers all of it into one addressable filesystem, reached through viking:// URIs, where memory, documents and skills are directories an agent can browse rather than an opaque index it can only query.
The structure is the interesting part. Context is layered, so a resource carries an abstract, an overview and the full content at separate levels, and an agent can read the cheap summary first and descend only when it needs the detail. That keeps token cost down on retrieval, which is the practical failure of naive RAG at scale. It ships first-party plugins for Claude Code, Codex, OpenClaw and OpenCode, a provider built into Hermes, and an MCP server for Cursor, TRAE, Manus and anything else that speaks the protocol, so it slots underneath an agent you already run rather than asking you to switch. The open source edition is AGPLv3 with the full feature set and no licence keys, which is a genuine constraint if you intend to embed it in a commercial product, and worth checking with a lawyer before you build on it.
Run the open source server yourself with Docker or Helm on Linux, macOS or Windows, or connect to the managed service. Installing a harness plugin is a single script that wires the agent to a server endpoint and an API key, after which the agent captures and recalls context automatically rather than through explicit commands. Content is addressed as viking:// paths: user memory and preferences under a user prefix, project documents and codebases under resources, session context and agent skills alongside them, with per-peer isolation for multi-tenant use. Web pages and GitHub repositories can be ingested and become browsable directories with the same layered abstract, overview and full-content structure. Retrieval is recursive, so the agent walks the tree rather than getting a flat list of chunks, and snapshots let you pin a known state of the context.
- •Filesystem Addressing - Memory, resources and skills as viking:// paths an agent navigates, instead of an opaque vector index it can only search
- •Layered Context - Abstract, overview and full content at separate levels so an agent reads the cheapest useful representation first
- •First-Party Harness Plugins - Claude Code, Codex, OpenClaw and OpenCode plugins plus MCP for Cursor, TRAE, Manus and others, with Python, LangChain and LangGraph SDKs
- •Recursive Retrieval - Walking the context tree rather than returning a flat set of chunks, with snapshots to pin a known state
- •Security in the Open Source Kernel - Envelope encryption with AES-256-GCM backed by local files, HashiCorp Vault or a KMS, per-account tenant isolation, and API-key or gateway-trusted auth
- •Observability Built In - Health and metrics endpoints, retrieval traces, and Prometheus and Grafana dashboards shipped with the open source edition
Teams running coding agents across many sessions or many people, where the same context keeps being rebuilt from scratch and the cost of that shows up in the token bill. It suits anyone already using Claude Code, Codex, OpenCode or OpenClaw who wants persistent memory underneath them without changing harness, and platform teams who need agent memory that is auditable and self-hosted rather than a vendor black box. Two things should give you pause. AGPLv3 means embedding it in a commercial product carries obligations, so the free edition is not free of consequences. And the managed service runs on Volcengine with the default endpoint in Beijing, with global hosting listed as coming soon, which makes it the wrong choice today for European or US teams with data residency requirements; those teams should self-host. The accuracy benchmarks published on the site are run in the project's own harness against competing memory layers, so read them as a claim rather than an independent result.
Pricing
Usage based, no flat monthly plan
- Open Source$0/mo
- OpenViking Context (Managed)Not listed
- Self-ManagedContact sales
From the vendor pricing page, 2026-09-19
