topodb-mcp-0.0.4 is not a library.

topodb-mcp

An MCP (Model Context Protocol) server exposing the TopoDB agent-memory engine over stdio. Point an MCP client at a .redb database file and it gets recall (get/find/search-text/search-vectors/traverse/access-stats/changes) and write (create memory, create entity, link, set-props, remove-node, close-edge, set-embedding, batch) tools backed by a scoped, temporal property graph — no separate database process, no network hop.

Status: v0 — read + write tools, including vector search and node/edge mutation. See Limitations.

Install

cargo install topodb-mcp

This installs the topodb-mcp binary to your Cargo bin directory (typically ~/.cargo/bin), which must be on PATH for the client configs below to find it by name.

CLI reference

topodb-mcp --db <path> [--scope <ulid|shared>] [--read-scopes <ulid|shared>[,...]]
           [--spec <path>] [--allow-unscoped-changes]

Flag	Required	Default	Meaning
`--db <path>`	yes	—	Path to the redb database file. A missing file is created on open; a missing parent directory is a startup error.
`--scope <ulid\|shared>`	no	`shared`	The default write scope: the one scope a create/link tool call is stamped with when it omits its own `scope` parameter. `"shared"` (case-insensitive) resolves to the shared scope; any other value is parsed as a ULID and resolves to that scope id. An invalid value is a startup error.
`--read-scopes <list>`	no	`--scope`'s value	The default read scope set: a comma-separated list of `"shared"`/scope ULIDs (whitespace around entries ignored) that a read tool call filters by when it omits both its own `scope` and `scopes` parameters. Reads filter by a set; `--scope` picks the single scope a write is stamped with — that asymmetry is why there are two flags instead of one. An empty list is a startup error (there is no unscoped read).
`--spec <path>`	no	inherit / built-in default	Path to a JSON file deserializing to `topodb::IndexSpec`, controlling which `(label, prop)` pairs are equality- or text-indexed, honored verbatim (may reindex an existing db). Omitted: an existing db inherits its own persisted spec (never reindexed or clobbered); a fresh db is created with the built-in default — equality index on `(Entity, name)`, text index on `(Memory, content)`, matching the labels/props the `create_entity`/`create_memory` tools write, so lookup and search work out of the box with no spec file. This mirrors `topodb-cli`, so a db either tool created is served identically by the other.
`--allow-unscoped-changes`	no	off	Bare toggle. `get_changes` is the one deliberately unscoped read tool — its op log spans every scope in the db — so it is rejected with `invalid_params` unless the server was started with this flag. Sync/consolidation hosts that legitimately need the whole log pass it.

Arg parsing is hand-rolled (five flags); there is no --help flag yet.

Tools

tools/list reports exactly 16 tools: db_info, six read tools (get_changes included), and nine write tools.

Tool	Params	Description
`db_info`	—	Report the open database's path, current op-log sequence number, the default WRITE scope applied to a create/link call that omits `scope`, and the default READ scope set applied to a read call that omits both `scope`/`scopes`. Call this first to confirm the server is wired to the expected database and read set, and to obtain `current_seq` as the anchor for `get_changes`.
`get_node`	`id` (string, required); `scope` (string, optional); `scopes` (string[], optional)	Fetch one node by its ULID. Call this when you already have a node id (from a previous search, traverse, or create) and need its current label and properties.
`find_by_prop`	`label`, `prop`, `value` (string/number/bool), `scope?`, `scopes?`	Exact-match lookup on an equality-indexed property (e.g. an Entity's name). Call this to resolve a known identifier to a node — NOT for fuzzy or full-text search; use `search_memories` for that. Errors if `(label, prop)` is not declared in the index spec.
`search_memories`	`query` (required), `k` (integer, default 10), `scope?`, `scopes?`	Full-text BM25 search over indexed text properties. Call this when looking for memories relevant to a topic or phrase. Returns up to `k` nodes ranked by relevance with scores.
`traverse`	`seed_id` (required), `max_hops` (integer, default 2), `direction` (enum: `out`/`in`/`both`, default `both`), `edge_types` (array of strings, optional), `scope?`, `scopes?`	Walk the graph outward from a seed node, following edges up to `max_hops`. Call this to gather the context AROUND something you already found — related entities, linked memories. Returns the subgraph (nodes + edges).
`access_stats`	`id` (required), `scope?`, `scopes?`	Read a node's access statistics (count, last-accessed timestamp). Call this when deciding what to consolidate or forget — e.g. finding stale memories. Reading stats does not itself count as an access.
`search_vectors`	`model` (string, required), `vector` (number array, required), `k` (integer, default 10), `candidates` (array of node ids, optional), `scope?`, `scopes?`	Cosine similarity search over embeddings stored under `model`. Call this when you have a host-computed query embedding and want nodes ranked by vector similarity rather than text relevance. `candidates` restricts scoring to a given node id set (e.g. narrow to a `traverse` result for hybrid recall). Errors if `k` is 0 or the vector is empty.
`get_changes`	`since_seq` (integer, required)	Replay the operation log from a sequence number (inclusive). Host-level primitive for consolidation/sync — the ONE unscoped read; the log spans all scopes. Returns ops with their seq numbers; on Compacted errors, re-anchor from current state. The `db_info` tool reports `current_seq`. Rejected with `invalid_params` unless the server was started with `--allow-unscoped-changes`.
`create_memory`	`content` (string, required), `props` (object, optional), `scope?`	Store a new memory. Call this when the user or task produces information worth remembering later. `content` becomes the full-text-searchable body; `props` holds structured metadata (strings/numbers/bools). Returns the new node's id — keep it if you plan to link this memory to entities.
`create_entity`	`name` (string, required), `props` (object, optional), `scope?`	Create an entity node (person, project, concept). Call this the FIRST time something is mentioned that memories should attach to; use `find_by_prop` first to check it doesn't already exist. `name` is equality-indexed for exact lookup.
`link`	`from_id`, `to_id`, `edge_type` (all required strings), `props` (object, optional), `valid_from` (integer ms, optional), `scope?`	Create a typed, time-aware edge between two existing nodes. Call this to connect a memory to the entities it concerns, or entities to each other (e.g. `'works_on'`). `edge_type` is free-form but be consistent — `traverse` can filter by it. Returns the edge id. Errors if either node doesn't exist.
`set_node_props`	`id` (string, required), `props` (object, required — a `null` value REMOVES that key)	Set or remove properties on an existing node. Errors if the node doesn't exist. Returns the committed seq.
`remove_node`	`id` (string, required)	Hard-delete a node and cascade-remove its incident edges. Call this to forget something entirely. Errors if the node doesn't exist. Returns the committed seq.
`close_edge`	`id` (string, required), `valid_to` (integer ms, optional — defaults to now)	Close an open edge, stamping its `valid_to`. Errors if the edge doesn't exist. Returns the committed seq.
`set_embedding`	`id` (string, required), `model` (string, required), `vector` (non-empty number array, required)	Attach a raw embedding vector (host-computed) to an existing node under `model`. Errors if the node doesn't exist, the vector is empty, or its dimension conflicts with the model's existing vectors. Returns the committed seq.
`submit_batch`	`commands` (array of command objects, required)	Submit a batch of high-level commands atomically — all commit or none. Each command's `op` matches a tool name (own field names, not always identical to that tool's param names — see the batch DSL). `#N` in an id field references the id produced by the Nth earlier command. Returns the produced ids in order (`null` for commands that create nothing).

Every scoped read tool accepts both scope (one scope) and scopes (an array of several, e.g. a project scope plus "shared") — a non-empty scopes wins over scope, which wins over the server's configured default read set (--read-scopes, or --scope alone). An explicitly empty scopes: [] is rejected (invalid_params) rather than treated as "read everything" — there is no unscoped read except get_changes, gated separately behind --allow-unscoped-changes. Every write tool accepts only scope (one scope) — see Scoping semantics for the full reads-filter-a-set-writes-stamp-one picture. scopes is not a write-tool param at all: every param struct rejects an unknown field rather than silently ignoring it, so passing scopes to a write tool is a clean tool error, not a quiet no-op. Engine errors and parse failures are returned as MCP tool errors carrying the engine's message — the server never panics on bad input.

Client configuration

Claude Code

Register a local stdio server with claude mcp add. Everything after the -- separator is passed to topodb-mcp untouched:

claude mcp add topodb --transport stdio -- topodb-mcp --db /path/to/agent.redb

Watch the --scope collision. claude mcp add has its own -s/--scope flag (registration scope: local/project/user — where the server config is stored). That is unrelated to topodb-mcp's own --scope flag (the default recall scope inside the database). Because ours comes after --, it's passed straight to the binary and there's no actual collision — but if you also want to set Claude Code's registration scope, put its --scope/-s before the --:
claude mcp add topodb --transport stdio --scope user -- topodb-mcp --db /path/to/agent.redb --scope shared

If you'd rather edit config directly, the equivalent stdio entry (project .mcp.json or ~/.claude.json) is:

{
  "mcpServers": {
    "topodb": {
      "command": "topodb-mcp",
      "args": ["--db", "/path/to/agent.redb"]
    }
  }
}

Claude Desktop

Add an entry to claude_desktop_config.json:

{
  "mcpServers": {
    "topodb": {
      "command": "topodb-mcp",
      "args": ["--db", "/path/to/agent.redb"]
    }
  }
}

Use an absolute path for --db. Claude Desktop spawns the server with its own environment, which may not include the same PATH as your shell — if topodb-mcp isn't found, replace "command": "topodb-mcp" with the absolute path to the installed binary (e.g. ~/.cargo/bin/topodb-mcp on macOS/Linux, %USERPROFILE%\.cargo\bin\topodb-mcp.exe on Windows).

Pi

One command — the @topodb/pi extension bundles everything (no Rust, no separate MCP adapter):

pi install npm:@topodb/pi

It registers a topodb tool that spawns this server for you. Config via env: TOPODB_DB (default .topodb/memory.redb), TOPODB_SCOPE (default shared).

Manual (any MCP server on Pi) — Pi has no built-in MCP client, so install an MCP client extension once, then point it at topodb-mcp:

pi install npm:pi-mcp-adapter

Then add topodb to the config that adapter reads (~/.pi/agent/mcp.json global, or .mcp.json project):

{
  "mcpServers": {
    "topodb": {
      "command": "npx",
      "args": ["-y", "@topodb/topodb-mcp", "--db", ".topodb/memory.redb"],
      "lifecycle": "lazy"
    }
  }
}

(pi-mcp-extension works too, but reads .pi/mcp.json instead of .mcp.json.)

Scoping semantics

TopoDB partitions nodes and edges into scopes — a shared scope plus any number of ULID-named scopes — so multiple agents or conversations can share one database file without stepping on each other's memories. Reads filter by a set of scopes; a write is stamped with exactly one. That asymmetry is the thing to get right: a read tool can gather results from several scopes at once (e.g. a private project scope plus shared), but a create/link tool call always picks exactly one scope for the node/edge it produces. topodb-mcp resolves scope as follows:

The server is started with two independent defaults:
- --scope (default shared) — the default write scope, used by create_memory, create_entity, and link when a call omits its own scope.
- --read-scopes (default: --scope's value alone) — the default read scope set, used by every scoped read tool when a call omits both scope and scopes. Comma-separated; an empty list is a startup error — there is no unscoped read.
Every scoped read tool call also accepts, per call: an optional scope (one scope), and an optional scopes (an array of several). Precedence: a non-empty scopes wins over scope, which wins over the server's default read set. An explicitly empty scopes: [] is rejected (invalid_params) rather than treated as "read everything".
Every write tool call accepts only a single optional scope, resolved against the server's default write scope (--scope), never against --read-scopes. link is the exception worth noting: its scope param determines which scope the edge itself lives in, independent of the scopes of the two nodes it connects — this is what lets an edge join nodes that live in a scope other than the server's default, e.g. an edge from a shared-scope entity to a private-scope memory. The batch DSL's link op takes the same scope field.
get_changes is the one deliberately unscoped tool: the operation log spans every scope, so a host can replay it for cross-scope consolidation or sync. There is no way to filter it by scope, and for that reason it is rejected with invalid_params unless the server was started with --allow-unscoped-changes.

If you want per-conversation isolation, start a separate topodb-mcp process per conversation with a distinct --scope <ulid> against the same --db file (or pass scope/scopes explicitly on each tool call from a single server instance).

v0 limitations

Bytes and DateTime prop values are unsupported over MCP. Only string, integer, float, and bool prop values round-trip through JSON; attempting to write or a stored node that contains a Bytes/DateTime prop is rejected/errors rather than silently coerced.
No HTTP/SSE transport. Only stdio, i.e. one client process per server process.

topodb-mcp 0.0.4