Sequential Thinking MCP Server

Sequential Thinking is a Model Context Protocol (MCP) server that operationalizes deliberate reasoning. It lets MCP clients create structured thinking sessions, capture sequential steps (analysis → hypothesis → verification → conclusion), branch into alternative lines of thought, and export or audit the full chain of reasoning.

Overview

• Written in Go 1.24 with the MCP Go SDK as its only direct dependency.
• Ships an in-memory session store, branch management, logical validation, and reasoning-quality heuristics.
• Provides eleven MCP tools, resource endpoints for sessions and templates, and three reasoning prompts.
• Supports stdio (default) and streamable HTTP transports, plus a minimal scratch-based Docker image published as bpradana/sequentialthinking.
• Backed by unit tests across handlers and an end-to-end MCP integration test.

Quick Start

Prerequisites

• Go 1.24 or newer
• Git (or another way to obtain the source)

Build and Run (stdio transport)

go mod download
go build -o sequentialthinking ./cmd/sequentialthinking
./sequentialthinking

The binary listens on stdio and is ready to be launched by MCP-aware clients.

Run as HTTP server

./sequentialthinking -http :8080

The -http flag enables streamable HTTP transport on the supplied address.

Docker (prebuilt image)

docker pull bpradana/sequentialthinking
docker run --rm -i bpradana/sequentialthinking

-i keeps stdin open so MCP clients can speak stdio with the containerised server.

Docker (build locally)

docker build -t sequentialthinking .
docker run --rm sequentialthinking

The multi-stage Dockerfile builds a static binary and runs it from a scratch image (UID 65532).

Integrating with MCP Clients

• MCP Inspector

  npm install -g @modelcontextprotocol/inspector
  mcp-inspector ./sequentialthinking
  

• Claude Desktop via Docker (macOS path shown)

  {
    "mcpServers": {
      "sequentialthinking": {
        "command": "docker",
        "args": [
          "run",
          "--rm",
          "-i",
          "bpradana/sequentialthinking"
        ]
      }
    }
  }
  

  Pull the image first with docker pull bpradana/sequentialthinking. Claude keeps the container alive while the MCP session is active and re-creates it on demand.

Kick off a Template Session

Once connected (Inspector, Claude, or another MCP client), call the sequentialthinking.start_from_template tool to seed a new session:

{
  "name": "sequentialthinking.start_from_template",
  "arguments": {
    "template": "scientific-method",
    "problem": "Why did the nightly build fail?",
    "context": {
      "repo": "widgets-service",
      "last_green_commit": "abc1234"
    }
  }
}

The server responds with the new session_id, populated steps from the requested template, and suggested next actions. Subsequent calls can extend or branch the reasoning as usual.

Tooling Surface

Resources and Prompts

• Resources

  • thinking://session/{session_id} – JSON snapshot of a single session.
  • thinking://sessions/list – JSON list of all sessions in memory.
  • thinking://template/{template_type} – JSON representation of thinking frameworks (scientific-method, five-whys, root-cause-analysis, decision-matrix, swot-analysis, pros-cons, first-principles, fishbone, pareto-analysis).

• Prompts

  • problem_breakdown – Structures complex problem decomposition.
  • critical_analysis – Guides evaluation of claims and evidence.
  • synthesis_prompt – Helps combine multiple insights into coherent output.

Reasoning Model Highlights

• Step Types: analysis, hypothesis, verification, conclusion. Parent references and connection tracking let clients build reasoning graphs.
• Branching: Any step can branch into alternative reasoning paths. Branch steps are numbered locally and remain tied to the originating step.
• Quality Scoring: Heuristic combines step-type diversity (30%), connection density (30%), and depth (40%). Scores stay in [0,1] and recompute after each update.
• Logical Validation: Detects unsupported conclusions and unverified hypotheses. Returns structured issues, suggestions, strength highlights, and an overall validity score.
• Pattern Detection: Identifies recurring reasoning patterns (hypothesis-verification, progressive refinement, branching synthesis) for analytics.
• Exports: Markdown, JSON, and text exporters include metadata, connections, branches, and suggested filenames.

Operational Notes

• Storage: In-memory (internal/thinking.MemoryStore) guarded by RW mutexes. Sessions are ephemeral and cleared on restart.
• Concurrency: Reads return deep clones to prevent external mutation; writes update timestamps, current step counters, and quality scores.
• Logging: Uses log/slog for initialization messages and mcp.LoggingTransport for stdio diagnostics.
• Completion Support: Autocomplete handler surfaces session IDs and template names for MCP completion requests.

Development Workflow

# Run unit and integration tests
go test ./...

# Format and lint
go fmt ./...
golangci-lint run    # optional, if installed

Project Layout

cmd/sequentialthinking   CLI entry point and flag parsing
internal/server          MCP server assembly, transports, integration tests
internal/handlers        Tool, resource, prompt handlers and helper logic
internal/thinking        Core domain types, in-memory store, exports, metrics

Release Tips

• Build with CGO_ENABLED=0 for static binaries (already used in Dockerfile).
• Bump the version in internal/server.New when publishing server changes.
• Use the provided Dockerfile as the canonical container build.

License

MIT License (use repository root LICENSE if present).