Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

OpenViking

Volcengine's self-hosted context server that stores an agent's documents, memories and skills as a browsable viking:// filesystem.

Reviewed by ReviewRamble editors · Updated

View on GitHub Open source

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

3.9/5

OpenViking makes agent memory something you can open and read, but its AGPL licence, Volcengine-first defaults and frequent breaking changes ask more of a team than Mem0 or Graphiti do.

What is OpenViking?

OpenViking is a server that holds everything an agent knows in one virtual filesystem under viking:// URIs: resources (documents, repositories, web pages), memories (preferences and experience extracted from conversations) and skills. Agents move through it with ls, tree, read, grep and find, and every processed directory carries a one-line abstract (L0) and an overview (L1) generated by a model, so an agent can read summaries before it opens full files (L2). Search can be scoped to a subtree rather than one flat vector pool.

Agents connect through an MCP endpoint served on the same port as the REST API, a memory plugin installer for Claude Code, Codex, Cursor, TRAE and OpenCode, native integrations for OpenClaw, Hermes Agent and pi, LangChain and Microsoft Agent Framework adapters, Python, Go and TypeScript SDKs, and the ov CLI. Committing a session archives the conversation and writes extracted memories as Markdown files you can open and edit.

The project comes from Volcano Engine (Volcengine), ByteDance's cloud business. The main server is AGPL-3.0; Volcengine also sells a hosted OpenViking Service and a licence-keyed Self-Managed edition with distributed deployment and support.

Is OpenViking worth it?

OpenViking's strongest idea is that memory should be inspectable. Instead of a vector store you query blind, it keeps documents, extracted memories and skills as files under viking:// paths, each directory topped with model-written summaries, and agents browse it with filesystem commands. The repository was created in January 2026 and has about 42,400 GitHub stars, 3,511 forks and nine releases between v0.5.0 on 9 October and v0.5.5 on 13 November 2026.

The evidence is vendor-run. The README reports LoCoMo long-conversation accuracy of 80 to 83% for OpenClaw, Hermes and Claude Code with OpenViking, against 24 to 57% on their native memory, with input tokens down 34.3 to 91.0%, and tau2-bench task success up 6.87 points in retail and 11.87 in airline. Those runs used version 0.3.22 with Doubao 2.0 Pro and Doubao-embedding-vision-251215, both Volcengine models, and no independent reproduction is published.

Models are your choice but not free. Embeddings default to a local GGUF model, bge-small-zh-v1.5, and the configuration guide lists 13 embedding providers including OpenAI, Azure, Gemini, Voyage, Jina, Ollama and LiteLLM. The vision-language model that writes summaries and extracts memories has no default; the setup wizard offers Volcengine, OpenAI, Codex sign-in, Kimi, GLM and Ollama, and other vendors go through LiteLLM. Every worked example in the docs uses Doubao models on Volcengine's Beijing endpoint, so expect to translate them. Each import and each committed session costs model calls.

Security defaults are sensible on one machine and need work beyond it. The server listens on 127.0.0.1:1933 and, with no root key set, runs in dev mode, where every request is treated as root; it refuses to start that way on a non-loopback address. The Docker image binds 0.0.0.0, so a root key is required there. API key, OIDC, LDAP and trusted-gateway modes are built in, and MCP supports OAuth 2.1 for Claude.ai and ChatGPT. Encryption at rest is off by default. In v0.5.5 the model gateway's default profile started exposing all tools, including eight that change data such as write, forget and set_acl, unless you list them in disabled_tools.

The licence decides most deployments. AGPL-3.0 obliges you to offer source to network users of a modified version you host, which matters if you patch the server inside a product. The README adds that rewriting the project wholesale and re-releasing it, or porting a substantial part to another language, needs Volcengine's written permission. Against Mem0 and Graphiti, both Apache-2.0, that is real friction, and the v0.5.5 release notes alone list 13 upgrade and migration items.

Where it earns its place

  • A developer using Claude Code and Codex on the same repositories

    Have both agents remember my preferences and past fixes without a cloud account

    Run openviking-server on localhost, then the curl installer adds hooks and MCP to each agent. Memories land as Markdown under viking://user/, so wrong ones can be edited or deleted. The software is free; the extraction model is billed per session by whichever provider you configure.

  • A platform team at a 200-person company building an internal assistant

    Give the assistant scoped search over our docs and per-user memory

    Deploy the Docker image with a root key and OIDC, create an account per team and user keys per person, and point the assistant at the REST API or Python SDK. Running it unmodified as a separate service is the common pattern; patching the server inside a hosted product brings AGPL source obligations.

  • A Chinese SaaS company already on Volcengine

    Add long-term memory to our agents without running another database

    Volcengine's hosted OpenViking Service runs the same server in Beijing; the first 50 files are free and plans and limits are in Volcengine's service documentation, not the repository.

Before you commit

Setup
`uv tool install openviking --upgrade && openviking-server init` (Python 3.10+ and uv), choosing an embedding model and a vision-language model; agents are then connected with `curl -fsSL https://openviking.ai/install | bash` (macOS or Linux, Node.js 18+). Docker image ghcr.io/volcengine/openviking and a Railway template are also published.
Integrations
MCP over Streamable HTTP at /mcp on port 1933, with a stdio proxy; Claude Code, Codex, Cursor, TRAE, OpenCode, OpenClaw, Hermes Agent, pi, DeerFlow; LangChain/LangGraph and Microsoft Agent Framework adapters; Python, Go and TypeScript SDKs; REST API; ov CLI; Web Studio at /studio.
Data and privacy
Self-hosted by default: content in a local workspace (./data) on the RAGFS file layer plus a vector index; pgvector, MySQL and Redis are options for shared deployments. Text is sent to the embedding and vision-language providers you configure. Encryption at rest and OpenTelemetry tracing are off by default. The hosted service runs in Volcengine's Beijing region; hosting outside China is planned on BytePlus.
Cost at scale
No licence fee for the AGPL server. The running cost is model calls: summaries for every imported directory and memory extraction on every committed session. Hosted and Self-Managed prices are not published in the repository.
Limits
1,541 open issues and pull requests; the v0.5.5 notes list 13 upgrade and migration items, several breaking; embedded vector storage does not support multiple processes, so shared deployments need a remote vector backend.

What’s good

  • Memories and documents stored as files you can read and edit
  • Directory-scoped search with summaries before full content
  • MCP, hooks, three SDKs and installers for most coding agents
  • Works with local, OpenAI, Gemini and other model providers

What’s not

  • AGPL-3.0, plus a README clause restricting rewrites and ports
  • Examples and benchmarks built around Volcengine's Doubao models
  • No auth and root access in dev mode until a key is set
  • Breaking configuration changes between point releases

Best for

Teams that want agent memory and document retrieval they can audit file by file, and can live with AGPL and a fast-moving server.

OpenViking alternatives

OpenViking: common questions

Is OpenViking worth it?
OpenViking makes agent memory something you can open and read, but its AGPL licence, Volcengine-first defaults and frequent breaking changes ask more of a team than Mem0 or Graphiti do. We rate it 3.9 out of 5.
How much does OpenViking cost?
Open source under AGPL-3.0, with no activation key. You pay for the embedding and vision-language model calls. Volcengine's hosted OpenViking Service has Personal and Enterprise plans, first 50 files free; prices are not in the README. Self-Managed edition by licence key, priced on request.
What are the downsides of OpenViking?
AGPL-3.0, plus a README clause restricting rewrites and ports; examples and benchmarks built around Volcengine's Doubao models; no auth and root access in dev mode until a key is set; breaking configuration changes between point releases.
Who is OpenViking best for?
Teams that want agent memory and document retrieval they can audit file by file, and can live with AGPL and a fast-moving server.
How much does OpenViking cost for a team or in production?
No licence fee for the AGPL server. The running cost is model calls: summaries for every imported directory and memory extraction on every committed session. Hosted and Self-Managed prices are not published in the repository.
Is OpenViking safe for company data?
Self-hosted by default: content in a local workspace (./data) on the RAGFS file layer plus a vector index; pgvector, MySQL and Redis are options for shared deployments. Text is sent to the embedding and vision-language providers you configure. Encryption at rest and OpenTelemetry tracing are off by default. The hosted service runs in Volcengine's Beijing region; hosting outside China is planned on BytePlus.
What are the best OpenViking alternatives?
The closest alternatives we have reviewed are Mem0, Graphiti, Cognee and claude-mem.

Answers checked Nov 18, 2026.