An MCP scratchpad that audits your agent’s tool plans

What it does

This MCP server is a structured scratchpad for model reasoning. It stores sequential thoughts per session, supports branching and revisions, and can validate that tool names recommended by the model appear in a user-supplied available_tools list. It also treats all input as untrusted and scrubs potential prompt-injection text before storage.

The interesting bit

The server deliberately does not discover tools or pick them for the model—it acts only as a validator and ledger. By requiring the caller to pass both available_tools and recommended_tools, it becomes a lightweight sanity check that prevents a model from hallucinating tool names into its plan. The branching and revision metadata let an agent backtrack or fork its reasoning without losing the audit trail.

Key highlights

• Rejects thoughts where recommended_tools contains names missing from available_tools
• Session-scoped history with branching, revisions, and a configurable per-session cap (MAX_HISTORY_SIZE)
• Prompt-injection scanning and redaction on all text fields, with security_warnings returned in affected calls
• Ships with a sequential-thinking-guidance prompt that tells the model to use the server as scratchpad, not as an external reasoning engine
• Adapted from the official MCP Sequential Thinking Server with added tool-plan validation

Caveats

• The repository description claims it “provides recommendations” for tools, but the README explicitly states it does not discover or choose tools for the model; it only validates names you supply
• The README warns against using it for trivial requests because it adds measurable overhead
• The security redaction is described as defensive filtering, not a guarantee against adversarial input

Verdict

Useful if you need inspectable, branching agent reasoning with a basic name-check on tool plans. Not what you want if you are looking for automatic tool discovery or an active recommendation engine.