VectorCode stalls as local MCP options multiply

The real story here is that “code indexing” is fragmenting into two camps: lightweight local filesystem tooling and heavier semantic/graph indexers. For teams that want everything on-prem, the maintenance cadence of the project matters almost as much as the indexing strategy.

• –VectorCode still looks like a capable local-first project, but its recent activity is a legitimate concern if you want a toolchain you can bet on long term
• –The strongest local setups seem to combine an MCP filesystem server with a local embedding model and vector store, rather than relying on a single all-in-one indexer
• –Newer projects like codebase-memory-mcp are pushing beyond embeddings into structural graphs, which is a better fit for dependency tracing and blast-radius analysis
• –The cloud-first smell some people dislike usually comes from auth, telemetry, or hosted backends, not MCP itself
• –This is less about “best embeddings” and more about whether the agent can traverse code like a developer would: grep, tree, imports, callers, and dependencies