ultrathink

UltraThink MCP Server is a Python-based MCP server that provides a structured, reflective thinking tool for solving problems. It supports sequential thinking with optional confidence scoring, auto-assigned thought numbering, and multi-session management. The server exposes a single tool called ultrathink, which accepts a series of thought steps and tracks how many steps are needed, how confident you are in each step, and how the reasoning branches or revises over time. You can run multiple concurrent thinking sessions identified by session IDs and review outcomes, uncertainty notes, and assumptions as the session progresses.

To use it, install and run the server via uvx, then interact with the ultrathink tool by sending thought payloads that describe each reasoning step. You can provide required fields like thought and total_thoughts, and optionally specify confidence, session_id, branching, revisions, and outcomes. The tool returns a structured JSON response including thought_number, total_thoughts, next_thought_needed, and a record of assumptions and risks, enabling you to trace the entire reasoning process.

Prerequisites:

• Python 3.12+ installation
• uvx (UltraVycle) tool installed (used to run Python MCP servers)

Installation steps:

1. Ensure Python 3.12+ is installed on your system. Verify with: python3 --version

2. Install or verify uvx is installed. If you don’t have uvx, install it using your preferred method per the uvx project (e.g., via pipx or your system package manager).

3. Run Ultrathink directly from GitHub without full installation: uvx --from git+https://github.com/husniadil/ultrathink ultrathink

4. Alternatively, clone the repository locally and use uvx in development mode: git clone https://github.com/husniadil/ultrathink.git cd ultrathink uv sync

5. (Optional) Run tests or example clients if you want to verify behavior using the provided tooling.

6. When ready, configure your MCP client to call the ultrathink tool as described in the usage section.

Tips and notes:

• The ultrathink tool supports session management via session_id, allowing multiple thinking streams to run concurrently.
• Confidence is a floating point between 0.0 and 1.0; use it to communicate how certain a given thought is.
• You can branch thoughts from a previous step using branch_from_thought and branch_id to explore alternative reasoning paths.
• There are many optional fields (uncertainty_notes, outcome, assumptions, depends_on_assumptions, invalidates_assumptions) to capture the full context of your thinking.
• Ensure your environment variables (if used by your deployment) are set appropriately, such as any API keys or internal service endpoints the reasoning pipeline depends on.
• If you plan to run the server in production, consider enabling multi-session support and configuring logging for better observability.