MCP SERVERLet agents read your whole Logseq graph but write only inside their own namespace, with every write logged to today's journal.
LABSCO SUMMARY
Two design choices do the work: agents write only under their own prefix, and each successful write appends a line to today's journal, so reviewing what they did is part of a note you already open. The guide tool and the named-query mechanism are the other half — they hand a model verified Datalog instead of letting it compose queries by hand.
A server for the file/Markdown version of Logseq, reached through Logseq's local HTTP API. Reads span the graph; writes are confined to one namespace prefix.
Logseq running with its HTTP APIs server enabled (Settings → Features, then start it from the plug menu) and an authorization token from those settings, passed as LOGSEQ_API_TOKEN. Installed from PyPI as mcp-server-logseq and launched with uvx over stdio; LOGSEQ_API_URL points at Logseq, default http://localhost:12315. A streamable-http transport is also available for remote clients and requires its own bearer secret, LOGSEQ_MCP_HTTP_TOKEN — the server refuses to start in that mode without one. A config file is optional; the defaults are read-mostly.
One command plus a key — uvx mcp-server-logseq, then supply credentials
USE CASES
Search block content across the graph, then read the pages worth reading in full with their block references resolved.
Pull task blocks by marker, priority or tag, then flip one to a new status without touching its text.
The agent writes a page under its own prefix and sets its properties; the write shows up as a line in today's journal.
Read the block, then replace its content by passing the exact current text — the edit is rejected if it has changed underneath you.
Read the graph's query guide first, then run the Datalog, or call one of the named queries kept in config.
List page names under a prefix to discover the child pages, then read the ones you need.
Counted once with only startup setup. Other versions, modes or optional credentials may expose a different list.
create_taskCreate a structured task. This is the ONLY way to make a task — write_note
rejects task-marker content. The block lives in the agent namespace, carries a
#task tag and a link to the executor agent, and optionally links a project,
a detail page, and a blocking block.
custom_queryRun a named custom query from the server config.
datascript_queryRun a raw Datalog query against the graph (advanced).
edit_blockReplace a block's content, confined to the agent namespace.
Read the block first: you must pass its exact current content as old_content. The edit is rejected if it doesn't match (you didn't read it, or it changed) — this enforces read-before-write and guards a concurrent edit. The block must live under the agent namespace. To change a task's status use set_task_status.
find_tasksFind task blocks by marker, tag, priority, page or agent-ownership scope.
get_logseq_guideReturn the authoritative guide for querying and writing this Logseq graph.
Read this once before using the other tools — it covers the verified Datalog gotchas (lowercase names, prefix descendants, marker/journal-day types, tags vs refs, read/write scoping) so you don't have to rediscover them.
list_custom_queriesList the named custom queries available from the server config.
list_pagesList page names under a namespace prefix (structure, not block content).
Namespace children are separate pages, so read_page on a parent shows none of them; use this to discover them. Matches the full lowercased `:block/name`.
read_blockRead a single block (and its children) with references resolved.
read_pageRead a page as a normalized block tree with references resolved.
searchFull-text search across block content in the graph.
set_page_propertiesSet or remove properties on an agent-namespace page.
set_task_statusChange only a task block's status marker (text untouched).
write_noteCreate or update a page in the agent's namespace.
BEFORE YOU INSTALL
14 tools when we connected. Each config query with register_as_tool = true adds a query_<name> tool, so the count can grow. set_task_status is gated by [tasks].allow_status_change.
Both documented clients run the server on your machine.
CONTEXT COST
Rated by the size of the 14 tool definitions that ride along on every turn, not by how many there are
INSTALLATION
uvx mcp-server-logseqclaude mcp add logseq --scope user --env LOGSEQ_API_TOKEN=<YOUR_TOKEN> -- uvx mcp-server-logseq{
"mcpServers": {
"logseq": {
"command": "uvx",
"args": [
"mcp-server-logseq"
],
"env": {
"LOGSEQ_API_TOKEN": "<YOUR_TOKEN>",
"LOGSEQ_API_URL": "http://127.0.0.1:12315"
}
}
}
}MCP servers
basicmachines-co
mohamadalhusseinie
JovanHsu
ukkit