The official sequential-thinking server holds the chain in memory for the length of one process. This one writes it down — so a session survives a restart, can be exported and handed to someone else, and can be analyzed for which stages you actually spent your thinking in.
A server for working through a problem in numbered steps, where each thought is categorized into a stage, stored with its metadata, and analyzed for its relationship to the others. Sessions persist on disk rather than living only in the conversation.
- `process_thought` — records a thought with its number, the expected total, whether more are needed, and a stage from Problem Definition, Research, Analysis, Synthesis or Conclusion. Optional tags, axioms used and assumptions challenged travel with it
- Revisions and branches on the same tool: mark a thought as revising an earlier one, or fork an alternative line of reasoning from a chosen thought under a named branch
- `generate_summary` — the whole process at a glance: total thoughts, the count per stage, a timeline marking revisions and branches, and per-branch statistics
- `clear_history` to reset, and `export_session` / `import_session` to write a session to JSON and read it back, including sessions exported by earlier versions
- Persistence as an append-only JSONL log where each call appends a single fsynced line, so the file doubles as an audit trail and a truncated final line from an interrupted write is recovered automatically
Python 3.10 or higher and the `uv` package manager. The shortest path is `uvx mcp-sequential-thinking` with no install step, or `pip install mcp-sequential-thinking` for a permanent one. Configuration examples are given for Claude Desktop, Cursor, VS Code, Zed, Claude Code, Windsurf and Gemini CLI, all running the published package through `uvx`. Sessions live at `~/.mcp_sequential_thinking/current_session.jsonl`, and `MCP_STORAGE_DIR` moves that directory; exports are confined to an `exports/` subdirectory beneath it.
One command — uvx mcp-sequential-thinking