think
Breaks complex tasks into manageable steps, tracks progress, and recommends tools for agentic problem-solving. Logs each thought in a thread to ensure clarity, self-correction, and iterative refinement for efficient task execution.
Instructions
Logs a single step in a thought process for agentic problem-solving. Supports thread following, step-tracking, self-correction, and tool recommendations. For each new user message, begin a new thought thread and log each thought after each completed step.
Key functionalities
Args: thread_purpose: A concise, high-level objective or thematic identifier for the current thought thread. Essential for organizing complex problem-solving trajectories. thought: The detailed, atomic unit of reasoning or action taken by the AI agent at the current step. This forms the core of the agent's internal monologue. thought_index: A monotonically increasing integer representing the sequence of thoughts within a specific thread_purpose. Crucial for chronological tracking and revision targeting. tool_recommendation: A precise actionable suggestion for the next tool to be invoked, omitted if no tool is needed, directly following the current thought. left_to_be_done: A flexible forward-looking statement outlining the next steps or sub-goals to be completed within the current thread_purpose. Supports multi-step planning and progress tracking. Omitted if no further action is needed.
Returns: A confirmation that the thought has been logged.
Example of thought process
user: "I keep hearing about central banks, but I don't understand what they are and how they work."
think(thread_purpose="Central banks explained", thought="Requires information about central banks and how they work. Consider using <named_tool> tool.", thought_index=1, tool_recommendation="<named_tool>", left_to_be_done="Summarize the findings and create an exhaustive graph representation")
call <named_tool>
think(thread_purpose="Central banks explained", thought="Summary of the findings is clear and exhaustive, I have enough information. Must create the graph with <named_tool>.", thought_index=2, tool_recommendation="<named_tool>", left_to_be_done="Send summary and graph to the user")
call <named_tool>
final: respond with summary and graph (no need to call think since left_to_be_done is a simple final step)
Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| left_to_be_done | No | None | |
| thought | Yes | ||
| thought_index | Yes | ||
| thread_purpose | Yes | ||
| tool_recommendation | No | None |
Input Schema (JSON Schema)
Implementation Reference
- sequential_thinking/mcp_server.py:11-63 (handler)The core handler function for the 'think' tool. It is decorated with @mcp.tool() for registration and implements the logging of sequential thoughts, tool recommendations, and task planning. The function signature defines the input schema.@mcp.tool() def think( thread_purpose: str, thought: str, thought_index: int, tool_recommendation: str | None = "None", left_to_be_done: str | None = "None", ) -> str: """Logs a thought step for agentic problem-solving, tracking reasoning, tools, and future plans. Start a new thread for each user message. # Capabilities - Workflow Orchestration: Breaks complex tasks into manageable steps. - Iterative Refinement: Self-corrects based on new info or errors. - Tool Recommendation: Suggests specific tools for the next action. - Forward Planning: Tracks remaining tasks via `left_to_be_done`. Args: thread_purpose: Short objective/topic for the thread. thought: Current reasoning or action description. thought_index: Monotonically increasing step number (1, 2, 3...). tool_recommendation: Tool to call next, or 'None'. left_to_be_done: Remaining steps/sub-goals, or 'None'. Returns: Confirmation of log. # Example 1) User: "I keep hearing about central banks, but I don't understand what they are and how they work." 2) think( thread_purpose="Explain Central Banks", thought="The user needs a comprehensive explanation of central banks. I need to identify their core definition, key roles (monetary policy, financial stability, currency issuance), and operational mechanisms (interest rates, reserves). I should search for a structured overview to ensure I don't miss major aspects like the Federal Reserve or ECB as examples.", thought_index=1, tool_recommendation="search_web", left_to_be_done="1. Search for 'how central banks work' and key functions. 2. Synthesize findings into a clear explanation. 3. Create a visual graph of the banking system structure if possible. 4. Present final answer." ) 3) call search_web(query="how central banks work and their main functions") 4) think( thread_purpose="Explain Central Banks", thought="Search results clarify that central banks manage currency stability, control inflation via interest rates, and act as lenders of last resort. I have enough textual information. Now, to make this easier to understand, I should create a graph representing the flow of money and influence between the central bank, commercial banks, and the economy.", thought_index=2, tool_recommendation="create_graph", left_to_be_done="Create a graph showing Central Bank -> Commercial Banks -> Public/Economy relations." ) 5) call create_graph(data=...) 6) Final Response: "Here is an explanation of central banks..." (Task complete, no further think call needed). """ log = f"Thread purpose: {thread_purpose}\nThought {thought_index} logged." if tool_recommendation and tool_recommendation.lower() != "none": log += f" Recommended tool: {tool_recommendation}." extra_log = f"{log}\nThought: {thought}" if left_to_be_done and left_to_be_done.lower() != "none": extra_log += f"\nNext: {left_to_be_done}" logger.info(extra_log) return log
- sequential_thinking/mcp_server.py:8-8 (registration)Creation of the FastMCP server instance where tools like 'think' are registered.mcp: FastMCP[Any] = FastMCP("Sequential Thinking")
- Input schema defined by the function parameters of the 'think' tool.def think( thread_purpose: str, thought: str, thought_index: int, tool_recommendation: str | None = "None", left_to_be_done: str | None = "None", ) -> str: