mcp-sequentialthinking-tools
mcp-sequentialthinking-tools is a lightweight MCP server that records a model's step-by-step reasoning as a branching, revisable history, and validates tool-plan recommendations against a supplied tool list, without discovering or choosing tools itself.
This server acts as a scratchpad for sequential reasoning: it stores one thought per call, along with its step number, an estimate of the total steps, and whether another thought is needed, and supports revising an earlier thought or branching a plan into an alternative path. It does not discover a client's other MCP tools or choose which one to use; if a caller supplies a list of available tools and its own recommended tools for the step, the server only checks that the recommended names exist in the available list and stores them.
It treats all of the thought text, tool descriptions, rationales and remaining-step text it receives as untrusted input, scanning for and redacting prompt-injection-like content before storing or returning it, and flags any call where a redaction occurred; this is described as defensive filtering, not a guarantee against all adversarial text. Stored history can be retrieved or cleared per session or across all sessions.
It is written in TypeScript and licensed under MIT.