Feedback on the Workflowy MCP (used daily with Claude as part of a multi-tool PKM setup: Workflowy, Roam, Tana). Probably re-stating other posts - but here are my list so far.
What works well: tree_seek is a strong filter engine (regex, mentions, time filters, subtree scoping, child/deep/match counts per hit), tree_patch batching is the best write model I’ve seen across PKM MCPs, and wait_any_match + leases are unique and valuable for agent workflows.
Suggested improvements, in priority order:
1. Attachment access. attachments_pull writes to ~/Library/Application Support/WorkFlowy/attachments/, which is sandboxed and unreadable for the agent, and files are named by node ID with no extension. Please allow a configurable cache path, include filename + MIME type in the result, and ideally return content directly (extracted PDF text with page numbers, images inline up to a size cap).
2. Read a full subtree in one call. A read_node/export tool that returns a node and its descendants as markdown with depth control and untruncated text (including notes) would remove many seek round-trips.
3. Semantic search. Opt-in embedding search over node text and notes (Roam ships this) is the single biggest gap for retrieval across a large outline – keyword/regex finds what you ask for, not what is related.
4. References and backlinks. A query for “all nodes that mention/mirror/link to node X” to treat Workflowy as a graph, not only a tree.
5. Query composition and scope. Explicit AND/OR/NOT across filter conditions, first-class #tag/@tag filters, and clear documentation of whether text filters cover notes as well as names.
6. Hosted/remote MCP option. Today the MCP requires the desktop app to be running; a remote endpoint (as Roam and Tana offer) would enable scheduled and mobile agent workflows.
7. Agent guidelines node. A designated node the MCP exposes as user conventions for AI agents (like Roam’s “roam/agent guidelines”), so rules travel with the data.
8. Schema clarity. Inline the $defs (tree_item_target, screencap_area, etc.) in tool descriptions with examples – agents currently have to guess target formats.
Hope this gives inspiration to continue the work with the MCP !