OpenViking

• Source: https://github.com/volcengine/OpenViking
• License: AGPL-3
• Status: RIP, replaced by qmd

Retired 2026-11. Everything the stack used it for, searching the workspace, qmd does on the CPU with no embedding model, no API key and no backup, matching exact words as well as meaning and fresher (60 seconds, not 15 minutes). Its own extras, a memory agents write to and a web UI, were never used, and a memory kept outside your Markdown files is against how the stack keeps knowledge. What follows is the page as it was.

Context database for AI agents: a viking:// filesystem of memory, resources and skills that persists between sessions.

Containers

Container	Role
openviking	REST API, MCP server, and Web Studio (port 1933)

The knowledgebase indexer cron runs in the main automation daemon, and the backup of openviking_data in automation-backup — OpenViking ships no container of its own for either.

Access

• Containers: http://openviking:1933 on the exist network
• Browser (Web Studio): https://openviking.EXIST_DOMAIN
• Agents: pi connects to it as an MCP server

Enable

EXIST_IS_AI_OPENVIKING=true

Then ./existential.sh && docker compose up -d from the repo root.

What it indexes

The workspace (volumes/workspace_data) — the same tree Nextcloud serves and code-server mounts at /workspace. Everything you actually work on is searchable without a second knowledgebase directory to keep in step.

The openviking_index_dir machine uploads it into viking://resources/workspace every 15 minutes. It is incremental: unchanged files are skipped, changed files replace their old copy, and files you delete on disk leave the index on the next run. A first pass over a large tree takes a while — each file is embedded on the way in.

The directory is mounted on the automation daemon, not on openviking. OpenViking's HTTP and MCP APIs both refuse host filesystem paths outright, so content only ever reaches it by upload — which is also why there is no watched-directory setup.

workspace/ai/ — where the agent automations write — is indexed like everything else, so an agent can find and build on what an earlier run produced. Writing there starts nothing unless a file handler matches it, so keep handler patterns off ai/ and that output never triggers more runs.

What it stores

Volume	Tier	Contents
openviking_data	2 — local, backed up	Vector store and workspace. Embedded DB, never NFS.

The upload manifest that keeps unchanged files from being re-embedded lives in the decree daemon's own decree_data volume (/data/openviking-index), not in a volume of its own. It is a cache: losing it costs one full re-index, not any content.