Every request takes the same route — no adaptive shortcut that makes results hard to compare — and the parallel specialists mean six angles land before anything is synthesised. The project states the cost plainly: roughly 5-10x the tokens per call. Treat it as the tool for the decision that deserves it, not the default reasoning path.
A sequential-thinking server backed by a team of agents. Every thought runs the same fixed path: an initial synthesis, six specialists in parallel, then a final synthesis.
- One tool, sequentialthinking, taking a thought, its number, the planned total, and flags for revision, branching and extending
- Six specialists running simultaneously — Factual, Emotional, Critical, Optimistic, Creative and Meta-cognitive — each with its own thinking direction and time budget
- A synthesis agent that runs twice, at the start and the end, on the enhanced model
- A structured result carrying should_continue, a stop_reason, the recommended next thought number and suggested arguments for the next call
- Revision and branching: mark a step as a revision when an earlier conclusion was wrong, or branch from a prior step with a branch id
- Optional web research through ExaTools, attached to every agent except synthesis
- Six providers supported, each with an enhanced and a standard model: DeepSeek by default, plus Groq, OpenRouter, GitHub Models, Anthropic and Ollama
Python 3.10+, uv or pip, and an API key for one of the supported providers — LLM_PROVIDER plus that provider's key, for example DEEPSEEK_API_KEY. EXA_API_KEY is optional and turns on research; without it the system runs on reasoning alone. Install from the repository, then add the mcp-server-mas-sequential-thinking command to your client with the environment set. Expect roughly 5-10x the tokens of a single-agent tool, since every call invokes several agents.
One command plus a key — uv pip install ., then supply credentials