// quick start
From zero to a first answer.
Requirements: Node.js 24+ and a git repository. You need some history to read — from smart-workbench’s ledger, from the plugin’s own log, or both.
Give it history to read
Install smart-workbench and turn its decision ledger on — or let the plugin record on its own with /vhdb:on record.
- >
/plugin install smartworkbench --marketplace nathak/smart-workbench - >
/swb ledger on
Add the VHDB plugin
The engine (engine/vhdb.wasm) is bundled; nothing to build.
- >
git clone https://github.com/nathak/vhdb-plugin ~/vhdb-plugin - >
claude --plugin-dir ~/vhdb-plugin
Search, then let Claude search
Try a query by hand, then turn on recording and injection.
- >
/vhdb:search token refresh kind:decision since:30d - >
/vhdb:on
Running a team server
Ubuntu 22.04+ / Debian 12+. The default config (auth = "off") accepts no accounts and listens on loopback only; set up Google login before sharing it.
- >
sudo apt install ./vhdb-server_0.1.0-1_amd64.deb - >
curl -s http://127.0.0.1:7600/healthz # {"ok":true}
// honest limits
What it doesn’t do yet.
- The plugin uses the local WASM engine only. Connecting it to the team server or daemon, and remote upload (
POST /v1/ingest), are not implemented. - No cross-project search: one token reaches one project, and a cross-project search would break that rule.
- Semantic search uses a hash-based stand-in; no real embedding model is connected.
- The log structure means individual history can’t be deleted. After a leak or an offboarding, the directory is rebuilt.
- Server login sessions are in memory only; a restart logs everyone out.
- Times are shown in UTC. Line-level history before a commit (edit hunks) is not recorded yet.
Give your sessions a memory.
MIT licensed. Source on GitHub.