dailydaniel/logseq-mcp

Turn your Logseq graph into memory and workspace for AI agents. A
Model Context Protocol server for
Logseq with safety-scoped writes, an audit trail in your
daily journal, and verified queries exposed as tools. Built on FastMCP (the
high-level API of the official mcp package).
> Targets the file/Markdown ("OG") version of Logseq — and plain-text files
> are part of why a graph makes good agent memory: git-syncable, greppable,
> durable, no lock-in. The newer DB (SQLite) version changed the underlying
> schema; some methods may behave differently there.
## Why
Agents need durable memory, and you already maintain one — your graph. The
missing piece is access you can trust: an agent should read broadly and write
usefully, but never touch what it shouldn't — and never do anything you can't
see. Three design choices make that possible:
Namespace-scoped writes. Agents write only under their own prefix
(byAgent/ by default), plus one deliberately narrow cross-namespace channel
that can change nothing but a task's TODO/DOING/DONE marker. Blacklisted
pages are hidden and redacted from every read.
An audit trail in your daily journal. Every successful write appends a
line like 22:30 [[byAgent]] wrote [[byAgent/readingList/...]] to today's
journal — reviewing your agents' work becomes part of a morning routine you
already have.
Verified queries as tools. Ship known-good Datalog from config as named
tools (queryweekplan, …), so agents don't compose datascript by hand and
cheaper models stay reliable.
## How I use it
I run a small fleet of Claude Code agents with this server on an always-on Mac
mini, against my live personal graph:
Nightly research. A link dropped into the reading list from the phone; at
night an agent claims it (status:: researching), reads the article — or
shallow-clones and reads the repo — writes a structured summary onto the page
and flips it to read.
Morning brief. At 08:30 a small model assembles a one-page dashboard —
what was read overnight, week-plan progress, current NOW/DOING tasks — and
sends a single push notification.
One journal for everyone. The human's tasks and the agents' audit lines
interleave in the same daily note:
The pages the researcher writes — properties, summary, relevance — link straight
into the rest of the graph:
mermaid
flowchart LR
A[AI agents] -- MCP tools --> S[logseq-mcp]
S -- HTTP API --> L[Logseq graph]
S -. audit line per write .-> J[daily journal]
Y((you)) --> L
Y -- morning review --> J
## Requirements
A running Logseq with the local HTTP API server enabled
(Settings → Features → HTTP APIs server, then start it from the 🔌 menu).
An authorization token created in the HTTP API server settings.
## Usage
### Claude Code
Local (stdio), token from the environment:
bash
claude mcp add logseq --scope user --env LOGSEQAPITOKEN= -- uvx mcp-server-logseq
Or point it at a remote instance over Streamable HTTP (how phone and remote
sessions reach a headless host — see Transports):
bash
claude mcp add logseq --scope user --transport http http:// :8000/mcp \
--header "Authorization: Bearer "
### Claude Desktop
json
"mcpServers": {
"logseq": {
"command": "uvx",
"args": ["mcp-server-logseq"],
"env": {