On your own machine

For one person and one collection there is a smaller MCP server that needs no vectrixdb serve, no sign-in and no keys. The client starts it as a command, it opens the collection from disk, and it offers search plus conversation memory: remember, recall, forget.

There is no sign-in on this server. To share a collection with other people, run vectrixdb serve with VECTRIXDB_MCP=1 instead: see Use your collections from an assistant over MCP.

Start it

pip install "vectrixdb[mcp]"
vectrixdb mcp --name notes --path ./data --mode hybrid

vectrixdb-mcp is the same server as a command of its own, which is what a client's config names. The default transport is stdio, which is what a desktop assistant starts as a command:

{
  "mcpServers": {
    "vectrixdb": {
      "command": "vectrixdb-mcp",
      "args": ["--name", "notes", "--path", "/abs/path/to/data"]
    }
  }
}

Give the path in full: the client starts the command from a folder of its own choosing.

Option	What it does	Default
--name	The collection.	default
--path	Where the database is.	./vectrixdb_data; for vectrixdb mcp, VECTRIXDB_PATH first
--mode	How it searches: dense, hybrid, ultimate or graph.	the library's default
--transport	stdio, sse or streamable-http.	stdio
--allow-writes	Over HTTP, also offer remember, feedback and forget.	off

vectrixdb mcp also takes --env-file, a settings file read first; what the environment already sets wins.

The tools

Tool	Arguments	What it does
search	query, limit, token_budget, mode	Search the collection.
recall	query, session, limit, token_budget	Conversation memories relevant to the query, weighted by recency and past feedback. session keeps to one conversation.
remember	text, session, role, pinned	Store a turn, or with pinned=true a fact that should always be in context.
feedback	id, outcome, correction	Grade a recalled memory: useful, dead_end or corrected. With a correction, the corrected text is stored and supersedes the old memory.
context	query, session, token_budget, recent_turns	A ready context block: pinned facts, the session's recent turns, then relevant memories, under one budget.
forget	session, older_than_days, ids, superseded, all_sessions	Delete memories for good: by ids, by age within a session, or only facts a correction already replaced. A call with none of these deletes nothing unless all_sessions=true. The one tool that deletes.

token_budget is 2000 unless given, and 1500 for context. recent_turns is 6. How the memory itself works is in Keep conversation memory without wasting tokens.

Over HTTP

--transport sse and --transport streamable-http serve it on this machine for other programs to reach. Over HTTP, remember, feedback and forget are left out unless --allow-writes is given, on either vectrixdb mcp or vectrixdb-mcp, since anything on the machine that reaches the port may call them. The server says so when it starts:

Serving search, recall and context only: --allow-writes adds remember, feedback and forget. For a team, run vectrixdb serve with VECTRIXDB_MCP=1.

Over stdio they are always there: the client that started the server is the only one talking to it.

Every answer says whether it was cut

Every tool that returns results takes a token_budget, and every answer starts with a line that says whether anything was cut:

[!] TRUNCATED: showing 3 of 9 results (~200-token budget). The answer may be among the 6 cut. Raise token_budget or narrow the query.

or, when nothing was:

[i] 9 results, ~410 of a 2000-token budget.

A tool result goes straight into the model's context, so a silently shortened list would read as a complete one and the model would reason from an absence that is not there. The header exists so that cannot happen. When a search fell back to vectors alone, a second line says so and why.

Use it from Python instead

The tool bodies are plain functions, so a program that already has a Vectrix can call them without the mcp package:

from vectrixdb import Vectrix
from vectrixdb.mcp_server import tool_context

db = Vectrix("memory", path="./data")
db.remember("we decided to ship on Friday", session="u42", role="user")

print(tool_context(db, "what did we decide", session="u42", token_budget=800))

build_server(db) returns the configured MCP server for a program that wants to mount it itself; build_server(db, writes=False) leaves out the tools that change the collection.

With tracing on, vectrixdb-mcp sends its spans to the OTLP collector itself, since nothing else in its process would.