Vibgrate AI Context is straightforward to run, but as with any MCP server, a few issues come up repeatedly. This reference collects the common problems with vg serve and how to resolve them.
It is for developers who have started the local context server and hit a snag connecting it or getting fresh answers.
The client does not see any Vibgrate tools
First confirm the server is running:
vg serve
Then confirm the client is wired to it. For supported assistants, re-running the installer rewrites the MCP configuration idempotently:
vg install
vg install supports Claude, Cursor, Windsurf, VS Code, Codex, and Gemini. For other MCP clients, verify their MCP configuration references the running server.
Answers are stale or miss recent changes
Served context reflects the code map you built. If you have changed code since, rebuild it:
vg build
Then check freshness — vg status compares the committed graph against your working tree and reports staleness:
vg status
The first run is slow
The first use downloads a small local model for semantic search. Precompute the index ahead of time so later runs are instant and offline:
vg embed
A library has no docs
If the assistant cannot find docs for a dependency, confirm it is resolvable:
vg lib
If it is missing, ensure the package is installed so Vibgrate can ingest it locally. Offline, rely on local ingestion rather than the hosted catalog.
The map looks empty
If structural answers come back empty, you may not have built a code map yet:
vg build
Verify what is being served
Use vg status to confirm node and edge counts and resolver state. Healthy counts and low staleness mean the server has real context to serve.
Related
- Starting the local context server with
vg serve - Offline and air-gapped AI context
- Connecting Vibgrate AI Context to a generic MCP client