Point an agent at it

The console's read surface is also an MCP server, so Claude Code — or any agent host that speaks MCP — can inspect a LanceDB database directly.

This costs nothing to run. There is no key of ours and no model in the loop: the intelligence is your agent's, and the evidence is the console's.

Connect it

claude mcp add lancescope -- \
  uv --directory /path/to/lancescope run python -m server.mcp_server

Then ask it something real:

what's in this database and what's wrong with it

It will come back with the tables, and with the unindexed vector column and what a search therefore costs — because those findings are already computed and it only has to read them.

Which database it reads

The same ladder the console climbs, resolved on every call: LANCE_ROOT, then the active saved connection, then the ingest directory if it holds tables. Switching connections in the console switches what the agent sees mid-session.

list_tables reports the root it resolved, and the server instructions ask the agent to name the database it is describing — you may have several.

To pin it to one database regardless of the console:

claude mcp add lancescope --env LANCE_ROOT=/path/to/tables -- \
  uv --directory /path/to/lancescope run python -m server.mcp_server

With nothing configured, every tool says so rather than guessing. An agent cannot tell a wrong answer from a right one, so the unconfigured state has to be unmistakable.

What it can and cannot do

Seven read tools, listed in the reference. Every one is declared read-only, and every one is the HTTP route called in process rather than a reimplementation — so the two surfaces cannot drift, and the guarantees hold in both.

It cannot materialise a blob column. read_rows has no expand parameter at all: the underlying route would refuse it, and not offering the argument means an agent cannot spend a turn discovering that. Reading every row of a table holding gigabytes of video costs kilobytes, however many times something asks.

It cannot write. Nothing under the console's routes writes, so nothing here does either.

It cannot spend your API budget. No summarise tool, no ask tool. The narrow set is deliberate.