Using the Sequential Thinking MCP Server to go from Generative to Agentic AI
What is the Sequential Thinking MCP Server from Anthropic?
The Sequential Thinking MCP Server is one of many Reference MCP Servers released by Anthropic to demonstrate the capabilities of their MCP protocol. This server in particular serves the purpose of providing structure to help augment a given AI’s thinking process. The Sequential Thinking server does not do any of it’s own “thinking” or decomposing of a problem. Instead it deterministically receives structured input from an AI, validates the data in the input, tracks the thought in an in-memory array for the session, and provides a pretty-printed version of the thought for inspection and debugging.
So what does this bring to the table?
The Sequential Thinking server gives the AI a workspace for thinking in a structured, deliberate, and auditable way. The structure encourages step-by-step reasoning and helps the AI revise and explore alternatives without having to be specifically asked by the user. It also provides valuable insight into the usually opaque process of how the AI reaches a conclusion, enabling debugging, auditing, and enhancing trust. This takes a lot of the guesswork out of the process of crafting prompts that consistently lead to specific desired outcomes.
The Sequential Thinking server also supports a more agentic style of operating from the AI that is self-directed, stateful, and goal oriented. Of course, for true agentic behavior the AI needs to have the ability to initiate actions and utilize tools beyond text generation, which is where the rest of the MCP server ecosystem comes into play.
How do you add the Sequential Thinking Server to Claude Desktop?
Prerequisites
First we’re going to explore the fastest way to get your hands dirty and start using the Sequential Thinking Server. What you will need to get started is:
Verify your installations with commands node -v and python --version , you may need to restart your terminal instance.
Open Claude for desktop and go through its onboarding process. The free account is sufficient for what we will be doing here.
Setup
Locate your claude_desktop_config.json file. If you are on a Mac it should be found at: /Users/<Your User>/Library/Application Support/Claude Open this config file in your preferred editor. For a new install it may only have empty curly brackets. You can replace that content with:
{
"mcpServers": {
"sequential-thinking": {
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-sequential-thinking"
]
}
}
}Save the file and restart your instance of Claude. In the chat window, you should see a plus sign next to a sliders icon. Click the sliders icon to see a list of installed MCP servers and verify that sequential-thinking is in that list. Click on sequential-thinking to verify that the slider is blue meaning it is turned on currently. If you ever want to disable an MPC server, you can return here to do so.
That’s it! Setup with the reference MCP servers is incredibly easy to get up and going.
The Fun Part: Using the Sequential Thinking Server
If you just want to see the server in action and how it changes your experience interacting with a generative AI (perhaps to use the transparency in it’s thinking process to better hone your prompts) then you’re all set. All you have to do is specifically direct Claude to use the Sequential Thinking tool within your prompt (otherwise it very likely will reason through your message without it). You don’t have to worry about capitalization, whether to use all one word or hyphenated or spaced – Claude understands which tool you mean regardless.
Head to head test: Claude with and without the sequential thinking server
To look for differences in reasoning I tried to find a challenging request that doesn’t have an easy answer and would require a good amount of reasoning and inference. Ultimately I went with the prompt: Please put together a comprehensive plan for making a simple house plant bioluminescent
Click to explore the conversation without Sequential Thinking
Click to explore the conversation with Sequential Thinking
- Claude without sequential thinking immediately returned a summary of three possible approaches, none of which were comprehensively planned and would still leave the user with the task of researching all three approaches to decide which to then develop a plan for. In this case, I would say it did not succeed at satisfying the request.
- Claude with sequential thinking invoked ( prompt:
Please use the sequential thinking tool to put together a comprehensive plan for making a simple house plant bioluminescent)spent a good amount of time analyzing the request and the problems, iterating 8 times to determine which approach would be the most feasible as well as most likely to satisfy the user’s requirements, and ultimately provided a single, in-depth plan for the most practical solution.
What would’ve been a lengthy, user-directed conversation that would be prone to errors if the user was asking about something outside their field of expertise (like a software developer looking to make bioluminescent houseplants) was shrunk down to a single iteration using the Sequential Thinking server. Pretty powerful from an end user perspective!
Combining the Sequential Thinking Server With Other MCP Servers to Go Further Along the Sliding Scale Between Generative and Agentic
Experiencing the upgrade in the responses provided by only utilizing the Sequential Thinking Server was very cool on its own. For our next trick, we’re going to utilize the google calendar integration with Claude Desktop in order to have Claude AI intelligently plan a great weekend in my hometown, Houston, Texas, factoring in my existing commitments for the weekend, the weather (spoiler, it’s hot in Houston), and the fact that I’m not a tourist.
To start, I asked Claude 3.7 Sonnet, ”Please help me plan a great weekend this weekend in Houston, Texas that fit into my already scheduled events. Please format your results as a guide with helpful links. Please factor in that I live here so I don't need to do the typical touristy things. Consider what things are popular and what's creating buzz for this weekend. Also consider what Houston is known for.”
Click for the chat without the multi-tool experience
Without invoking either the Sequential Thinking or Google Calendar tools, I got some good suggestions for events, but the AI went rouge on assumptions about my calendar without trying to use the paid integration, blocking of Friday and Saturday mornings for my previously scheduled events which were actually scheduled in the afternoons. Sunday was free but they AI blocked off Sunday evening as well. Not ideal!
By inserting the phrase “use the sequential thinking and google calendar tools to” into the prompt, the AI was able to formulate a plan of attack, actually retrieve my real schedule (including my all-important “Lunch with my cat”), research events from sites I recognize as having good non-touristy options, research the top options and the weather, and combine that into a high quality result.
Click for the multi-tool enabled chat’s Weekend Guide
It even gave me a very clear heat advisory warning, which was both accurate and important to the options.
It explained what was going on in town the way a human would, with solid context. With the Greek Fest it let me know it was the last day, which is good to know because I usually make it every year but had forgotten this year! It also very accurately pointed out that parking at the Greek Fest is a nightmare, and came to the same conclusion I did last year (which was to take a rideshare option in the future).
It organized suggestions based on convenience, noting things like “Since you’ll already be at POST for the market, explore the diverse food hall featuring local vendors and unique culinary offerings.” Everything it suggested was on-point, including the suggestion to see Drunk Shakespeare (Hilarious, I’ve been!) and getting dinner in Asiatown, which in Houston has an incredible plethora of delicious and unique options.
When using the paid Google Calendar integration, the chat itself can’t be directly shared, but here is a portion of its thought process:
So yes, the results are much better and do use multiple tools, but does that make them agentic?
Key elements of “agentic” behavior are autonomy, goal directed behavior, environmental interaction (using tools like web search, code execution, or calendar management) and external action capabilities, i.e. the ability to affect something beyond text/image/document generation. That last element is the sticking point that lands this experiment with the Semi-Agentic title. Had we implemented an MCP server for Claude to write google calendar events back to my calendar, this rubric would be satisfied. It’s important to note that there are a million different “rubrics” for defining agentic behavior and that its definition is still evolving. Merriam Webster still considers it slang.
Something to bear in mind is that being agentic is not the same thing as truly having agency. It just means that it mimics or behaves like it has agency, which is a deeply philosophical topic best discussed over delicious Sichuan noodles on a Saturday evening.
Stay tuned for more experiments and insights from P2 Labs!