OpenViking
Also known as: OpenViking Service, OpenViking Context, VikingBot, openviking.ai
OpenViking is an open source context database for AI agents from Volcano Engine, Volcengine for short. It keeps an agent's knowledge, memory and skills as one browsable filesystem, loads summaries first and full content only when needed, and turns each committed session into editable memory files. Hooks and plugins connect it to Claude Code, Codex, Cursor, OpenClaw and LangChain, and it runs self hosted or as a managed Volcengine service.
OpenViking is an open source context database for AI agents, published by Volcano Engine, also known as Volcengine, with its contributors under the AGPL 3.0 license. Its GitHub repository has more than 41,000 stars. In Volcengine's benchmarks with Doubao models, adding OpenViking raised LoCoMo long conversation accuracy from 24.20% to 82.08% for OpenClaw, from 33.38% to 82.86% for Hermes and from 57.21% to 80.32% for Claude Code, and cut input tokens by 34.3% to 91.0%.
Everything an agent knows lives in one virtual filesystem under viking:// addresses: resources, memories, skills and rules. Agents browse it with ls, tree, read, write and grep, and people can open the same directories to inspect or edit what the agent knows. Each item has a one sentence abstract, an overview and full content, and the agent loads the deeper levels only when it needs them. Semantic search can be limited to a project or a memory subtree, and a planning search draws its query from the session.
When a session is committed, OpenViking archives the conversation and extracts preferences, project decisions and reusable experience as Markdown memory files in the background, listing each addition and update in a memory diff. Later sessions recall those memories and cite their addresses. Rules, in beta, compile a team's own sources into mandatory requirements an agent can check its work against, and team management covers accounts, users and access control, with memory kept in each user's own space.
Integrations meet agents where they run. Memory plugins with lifecycle hooks capture and recall context automatically for Claude Code, Codex, Hermes and OpenClaw, a single command adds hooks, MCP tools, rules and skills to Cursor, LangChain and LangGraph can use it as a retriever or store, and any MCP client can point at its built in endpoint. VikingBot, a CLI, an SDK and an HTTP API complete the surface, and a manual evaluation kit gives test cases for retrieval and cross session memory. Deployments need an embedding model and a vision language model from Volcengine, OpenAI, Kimi, GLM or local Ollama.
Teams can self host the open source server with Docker or Helm, use Volcengine's managed OpenViking Service with the first 50 files free, or license the self managed edition for private cloud or offline use with official support. OpenViking suits developers and platform teams that want long term memory and organized context for coding agents and custom agents, especially those willing to run their own infrastructure. Teams that need published compliance attestations or a workflow engine will find a context layer to pair with other tools.
Vendor details
Canonical URL
https://www.openviking.ai/
Category
Agent infrastructure
Subcategory
Open source context and memory database for AI agents
Company status
first party product
Use cases & customers
Primary use cases
Target customers
Deployment options
Integrations
Hooks and plugins for Claude Code, Codex, Cursor, OpenClaw, Hermes, OpenCode and TRAE, LangChain and LangGraph, a built in MCP endpoint, an SDK and HTTP API, and model providers including Volcengine, OpenAI, Kimi, GLM and local Ollama.
In practice
A developer's Claude Code sessions are captured through hooks, and next week's session recalls the project's conventions without being told again.
A support agent built with LangGraph retrieves only the overview of a policy folder first and loads the full document when a case needs it.
A bank licenses the self managed edition to run agent memory inside its own network with no outside connection.
Sources & related URLs
Related / legacy domains
Research sources
Agentic Index coverage score
8.0 / 14 capabilities · 57%
| Integrations & Tool Calling | Partial |
|---|---|
|
OpenViking gives agents tools to list, read, write, search and grep its own context store, and plugs into Claude Code, Codex, Cursor, OpenClaw, Hermes, LangChain and any MCP client. Those tools act on the agent's knowledge and memory rather than on business systems, so taking action in outside applications stays with the agent framework that uses OpenViking alongside other tools. Sourcedocs.openviking.net/en/agent-integrations/01-overviewread 2026-11-20 |
|
| Workflow Orchestration | Not documented |
|
OpenViking supplies context and stores memory, and the agent that reads it decides what to do next. Session commits run memory extraction in the background, and context compilation turns sources into rules and summaries. Neither sequences tasks, branches on outcomes or coordinates several agents, which remains the job of the harness or framework connected to it. Sourcedocs.openviking.net/en/getting-started/01-introductionread 2026-11-20 |
|
| Knowledge Grounding & RAG | Full |
|
Resources, memories, skills and rules live in one persistent store under viking:// addresses, each with a one sentence abstract, an overview and full content that load in that order as needed. Semantic search can be limited to a project or memory subtree, a planning search draws queries from the session, and results point to sources an agent can open to check its conclusions. Sourcegithub.com/volcengine/openvikingread 2026-11-20 |
|
| Human Oversight & Guardrails | Not documented |
|
Memory extraction writes a diff of each addition and update that a person can read, and rules, in beta, compile mandatory requirements for an agent to check its work against. Extraction still commits on its own without waiting for approval. Approving what an agent does with the context it receives remains with the agent framework or tool calling OpenViking. Sourcedocs.openviking.net/en/getting-started/09-memoryread 2026-11-20 |
|
| Security, Identity & Governance | Partial |
|
Remote use requires an API key, and team management covers authentication, accounts, users and access control, with memory kept in each user's own space. Self hosted operators handle access control, backups and upgrades themselves, and the open source server runs under AGPL 3.0. No security attestation scoped to OpenViking or its managed service appears alongside those controls. Sourcedocs.openviking.net/en/getting-started/07-optionsread 2026-11-20 |
|
| Observability & Auditability | Partial |
|
Each session commit lists the memories it added and updated in a diff file, and recalled memories carry their addresses so a person can check which ones an answer used. The console and Studio show what the store holds at any time. A trace of each retrieval and the reasoning behind what was loaded is not described for teams that review agent behavior. Sourcedocs.openviking.net/en/getting-started/09-memoryread 2026-11-20 |
|
| Memory & State Persistence | Full |
|
Committing a session archives the conversation and extracts preferences, project decisions and reusable experience into Markdown memory files in the user's own memory space, with a diff of each addition and update. Later sessions recall them by address, and because memory is a filesystem, people can open, edit or remove any file. Saving and recall require the same service and user identity. Sourcedocs.openviking.net/en/getting-started/09-memoryread 2026-11-20 |
|
| Deployment & Data Residency | Full |
|
The open source server runs on a team's own machine, servers or cloud through Docker or Helm, Volcengine hosts a managed OpenViking Service, and a licensed self managed edition with official support covers private cloud and offline deployments on Kubernetes with VikingDB. Embedding and vision models can also run locally through Ollama, keeping context off outside services. Sourcedocs.openviking.net/en/getting-started/07-optionsread 2026-11-20 |
|
| Prebuilt Agents / Templates / Packs | Partial |
|
Ready made plugins connect OpenViking to Claude Code, Codex, Cursor, OpenClaw, Hermes, OpenCode and LangChain, and VikingBot ships as an agent that already uses the store. Skills and rules held in OpenViking are ones a team writes itself or compiles from its own sources. No library of packaged agents or industry context sets comes ready to adopt as is. Sourcedocs.openviking.net/en/agent-integrations/01-overviewread 2026-11-20 |
|
| Triggers & Channel Coverage | Partial |
|
Lifecycle hooks in Claude Code, Codex, Hermes and OpenClaw capture sessions and recall context automatically as an agent starts and ends work, and committing a session starts memory extraction in the background. Requests otherwise arrive from the CLI, VikingBot, the SDK, the HTTP API or MCP clients. OpenViking runs no schedules or event listeners that wake an agent itself. Sourcedocs.openviking.net/en/agent-integrations/01-overviewread 2026-11-20 |
|
| Model Flexibility & Routing | Full |
|
Every deployment chooses its own embedding model and vision language model during setup, with providers including Volcengine, OpenAI, Codex sign in, Kimi, GLM and local Ollama. Those models power indexing, summaries and memory extraction inside OpenViking, so the operator decides what runs where. Volcengine's published benchmarks used its Doubao models. Sourcegithub.com/volcengine/openvikingread 2026-11-20 |
|
| APIs / SDKs / MCP Extensibility | Full |
|
OpenViking exposes an SDK and HTTP API with documented conventions, a CLI, and a built in MCP endpoint that any MCP client can call for its tools, with API keys for remote use. A gateway serves compatible clients, and the full source is open under AGPL 3.0, so teams can extend the server itself or build their own integrations and plugins on top of it. Sourcedocs.openviking.net/en/getting-started/01-introductionread 2026-11-20 |
|
| Testing, Debugging & Optimization | Partial |
|
An evaluation kit gives test cases for retrieval and cross session memory with pass criteria, a downloadable sheet for recording results and guidance for tracing failures to import, indexing, reads, generation or memory. People run the cases and score them by hand. Volcengine's published LoCoMo and tau2 benchmark results describe the product rather than a team's own agents. Sourcedocs.openviking.net/en/getting-started/06-evaluateread 2026-11-20 |
|
| Browser / Computer-use | Not documented |
|
OpenViking stores and serves context through filesystem style tools, an API and MCP, and it imports the resources that teams add to it, such as documents and repositories. It does not operate a browser, a desktop session or a remote computer for an agent. Agents that need to use websites do so through their own tools while reading context from OpenViking. Sourcedocs.openviking.net/en/getting-started/01-introductionread 2026-11-20 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Pricing
Free and open source; managed service free for the first 50 files
Open source at no charge; managed service by files stored; self managed edition by license
Included quota
Volcengine's managed OpenViking Service includes the first 50 files free.
What is public
OpenViking is free and open source under AGPL 3.0 for self hosting. Volcengine's managed OpenViking Service includes the first 50 files free, and the self managed edition for private cloud or offline use requires a license key.
Billing mechanics
Self hosted teams pay for their own infrastructure and for the embedding and vision models they configure. The managed service connects with an API key from the Volcengine console.
Sales call required
Mixed (some tiers require a call)
Free / trial
Free and open source
Commercial notes
The licensed edition comes with official support and runs on Kubernetes with VikingDB.
Missing data
Volcengine sets managed service prices beyond the first 50 files in its console, and licenses for the self managed edition come through Volcengine.
Related vendors
- AgentOps — Agent observability and debugging platform: open source SDKs trace…
- AgentPhone — AgentPhone gives the AI agents a team builds their own US or…
- Agno — Python agent framework and AgentOS runtime (formerly Phidata) for…
- Airbyte — Airbyte moves data from operational systems into warehouses and…
- AIsa — Resource and payment gateway for AI agents, with one key to 110+…
- Albato — Albato is a no code integration platform that sells an embedded,…