MCP server
rekall mcp exposes your stores as Model Context Protocol tools, giving Claude Code, Claude Desktop, and Cursor millisecond neural-search inner-loop retrieval with one line of config.
The Model Context Protocol (MCP) is an open standard for
giving AI agents tools. rekall mcp runs a stdio MCP server that exposes your Rekall
stores as tools — so an agent gets millisecond, reasoning-grade retrieval in its inner loop
with one line of config.
The wedge
4 ms is fast enough to sit inside an agent's loop. Wire Rekall in as MCP tools and your
agent can query its memory hundreds of times per task — with a confidence signal to
branch on — without blowing the latency budget.
Start the server
rekall mcpIt ships with the CLI client and speaks MCP over stdio, so you normally don't run it by
hand — you point a client's config at it and the client launches it. It reads
REKALL_API_KEY (and optional REKALL_BASE_URL) from the environment, so the same server
works against the managed cloud or an enterprise self-hosted deployment.
Client configuration
One command:
claude mcp add rekall -- rekall mcpOr add it to your project's .mcp.json (checked in, shared with your team):
{
"mcpServers": {
"rekall": {
"command": "rekall",
"args": ["mcp"],
"env": {
"REKALL_API_KEY": "rk_live_…",
"REKALL_BASE_URL": "https://api.rekall.sh/v1"
}
}
}
}REKALL_BASE_URL is optional — it defaults to the managed cloud. Set it to your
enterprise deployment (e.g.
https://rekall.internal.acme.com/v1) to point the agent's tools at your own network.
Edit claude_desktop_config.json (Settings → Developer → Edit Config):
{
"mcpServers": {
"rekall": {
"command": "rekall",
"args": ["mcp"],
"env": {
"REKALL_API_KEY": "rk_live_…",
"REKALL_BASE_URL": "https://api.rekall.sh/v1"
}
}
}
}Restart Claude Desktop; the Rekall tools appear in the tools menu.
Add to ~/.cursor/mcp.json (global) or .cursor/mcp.json (per-project):
{
"mcpServers": {
"rekall": {
"command": "rekall",
"args": ["mcp"],
"env": {
"REKALL_API_KEY": "rk_live_…",
"REKALL_BASE_URL": "https://api.rekall.sh/v1"
}
}
}
}Tools
The server exposes three tools.
rekall_search
Run the reasoning loop over a store. Returns a SearchResult.
Prop
Type
{
"type": "object",
"required": ["query", "store"],
"properties": {
"query": { "type": "string" },
"store": { "type": "string" },
"max_hops": { "type": "integer", "minimum": 1, "maximum": 8, "default": 6 },
"max_results": { "type": "integer", "minimum": 1, "maximum": 50, "default": 5 },
"return_trace": { "type": "boolean", "default": true },
"filter": { "type": "object" }
}
}rekall_ingest
Write documents into a store (upsert). Lets an agent grow its memory within a task.
Prop
Type
{
"type": "object",
"required": ["store", "documents"],
"properties": {
"store": { "type": "string" },
"documents": {
"type": "array", "minItems": 1, "maxItems": 1000,
"items": {
"type": "object",
"required": ["id", "text"],
"properties": {
"id": { "type": "string" },
"text": { "type": "string" },
"metadata": { "type": "object" }
}
}
}
}
}rekall_list_stores
List the stores available to the configured project. No arguments.
{ "type": "object", "properties": {} }Using it in an agent
Once configured, the agent calls the tools like any other. A typical inner loop searches memory, acts, then writes the observation back so memory grows within the task — see the agent inner-loop memory recipe.
Next steps
Self-hosting (Enterprise)
Run the SKIM engine inside your own network — a licensed single binary plus a versioned model artifact. Cloud Run, Kubernetes, or VMs. The exact same API as the managed cloud.
CLI reference
Every rekall command and flag — login, ingest, search, stores, docs, mcp, keys — with realistic output examples. The CLI is a client for the Rekall API.