Sequential Thinking MAS MCP
An MCP server that runs every reasoning step of your agent past six Agno specialists and merges their views into one answer
- 305 ★ on the repository
- Last commit
Extended access
Why this level
- The text of every thought goes to the external model provider you chose
- Provider keys reach the server through environment variables
- With EXA_API_KEY set, the specialists reach out to the internet
Install
- Works in
Manual install
git clone https://github.com/FradSer/mcp-server-mas-sequential-thinking.git && cd mcp-server-mas-sequential-thinking && uv pip install .pip install . works instead of uv. Python 3.10 or newer is required.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The text of every thought goes to the external model provider you chose
- Provider keys reach the server through environment variables
- With EXA_API_KEY set, the specialists reach out to the internet
This is third-party code. Review the repository files before installing.
What it does
Sequential Thinking MAS adds one tool to an MCP client, sequentialthinking. The client sends one thought per call, and the server passes each through the full_exploration route built on the Agno framework. First a synthesis agent makes an initial summary. Then six specialists think in parallel: three check facts, intuition and weak spots, and three more look for benefits, new ideas and flaws in the reasoning itself. Finally synthesis assembles a single answer. A complexity analyzer records a score, the problem type and the thinking modes needed, for diagnostics. Specialists have different time budgets, from 30 seconds for the emotional one to 240 for the creative one. With EXA_API_KEY set, they can search the web through ExaTools. The server reply carries a should_continue flag, the next thought number and ready-made arguments for the next call. The tool is meant as a loop: the client keeps calling it until should_continue becomes false and sends a revision with isRevision=true for a weak step. Models come from the chosen provider, DeepSeek by default, and a local Ollama works too.
Who it is for. For people who solve hard problems step by step with an agent and want each step to pass through different viewpoints and merge into one conclusion.
Good fit when
- You need to weigh an architecture decision from several sides, facts, risks, benefits and alternatives, before writing code
- The agent must revisit a weak reasoning step, and being able to send a revision or a branch from an earlier step matters
- You need a debate over a contested idea, with one agent hunting for weaknesses and another proposing options
- You want to run the reasoning on your own set of models, including a local Ollama, with optional web search through Exa
Not a fit when
- The task is simple and one answer settles it
- You have no key for any supported provider and no local Ollama running: the server needs a model
- You want a light notebook for recording steps: Sequential Thinking MCP Server and mcp-sequentialthinking-tools in the catalog cover that
- Low latency and cost matter: synthesis runs twice and specialist time budgets reach four minutes
Example request
Use sequentialthinking to work through, step by step, whether we should move from a monolith to modules: facts, risks, benefits and alternativesMCP
- Transport
- stdio
- Authentication
- API key
Environment variables
LLM_PROVIDERrequired- Model provider: deepseek, groq, openrouter, github, anthropic or ollama
DEEPSEEK_API_KEYrequired, secret- Key of the chosen provider, here for DeepSeek; other providers use their own variable
EXA_API_KEYsecret- Turns on web search through ExaTools for the specialist agents
TEAM_MODE- Team orchestration mode: standard, route or coordinate
Limitations
The server calls an external model, so you need Python 3.10 or newer and a key for one of the supported providers. Only a local Ollama works without a key, and that setup does not depend on outside services. The Anthropic API is unavailable from Russia.
The strategy is fixed: every thought takes the full route, and the complexity score serves only diagnostics. So even a short question goes through double synthesis and all six specialists, and each of them spends model requests.
How to disable. Remove the sequential-thinking entry from your MCP client's configuration and restart the client. The package itself is removed with pip uninstall if you installed through pip or uv pip.
FAQ
How do I connect Sequential Thinking MAS to an MCP client?
Clone the repository and run uv pip install . or pip install . Then add a server with the command mcp-server-mas-sequential-thinking to the client config, with LLM_PROVIDER and your provider key, for example DEEPSEEK_API_KEY.
Which agents take part in processing a thought?
Six specialists: factual, emotional, critical, optimistic, creative and metacognitive. The synthesis agent runs twice, at the start and at the end.
Which model providers are supported?
DeepSeek by default with deepseek-chat, then Groq, OpenRouter, GitHub Models, Anthropic and Ollama. The enhanced model goes to the synthesis agent and the standard one to the specialists, and you can override both with variables like DEEPSEEK_ENHANCED_MODEL_ID.
Do I need an Exa key?
No. Web search through ExaTools turns on only when EXA_API_KEY is set. Without it the system reasons without touching the internet.
How should I call the sequentialthinking tool?
As a loop. After each reply read structuredContent.should_continue and repeat the call until it is false. Take next_thought_number and next_call_arguments for the next call, and fix a weak step with a revision using isRevision=true.
Discussion
No questions or reviews yet.
Related
MCP serversA local MCP server that keeps a staged journal of an agent's reasoning with persistent sessions, decisions and evidence
A lightweight MCP server for step-by-step reasoning: a scratchpad with history, branching and tool-plan validation
An MCP server for long-term agent memory: compresses dialogues into atomic facts, resolves pronouns and anchors relative time to absolute dates
An MCP server that inserts a human feedback step into agent tasks: a web or desktop window for text, images and prompts
An MCP server that turns a request into a step plan with dependencies and keeps context across agent sessions
A skill with 25 commands to audit and improve the prompts, tools and agent designs of a project in Cursor, Claude Code and Codex
Comparison with similar tools
Entries that share the main search query, and how they differ.
| Entry | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| Sequential Thinking MAS (this entry) | no | API key | 305 | |
| Sequential Thinking MCP Server | no | not required | 959 | a local staged journal of the agent's reasoning with persistent sessions, without a team of specialists |
| mcp-sequentialthinking-tools | no | not required | 586 | a lightweight scratchpad for step-by-step reasoning with history and branching |
README in short
The README opens with the route diagram and a table of specialists with time budgets. The tool's input and output format, the call contract and a provider table with variables follow. Installation goes through cloning and uv.