Best for
- Developers needing transparent, auditable AI reasoning
- Researchers decomposing complex problems into manageable subtasks
- Users of MCP-aware clients like Claude Code, Cursor, Windsurf
A
Atom of Thoughts MCP Server
Community
Structured reasoning for LLMs: decompose, track confidence, visualize, approve.
GitHub source repository ↗
54FMRS · D
This server enhances AI decision-making with structured reasoning, improving explainability. Installation is easy, supports multiple clients. Suitable for developers needing transparent reasoning, but be aware of external dependencies and trust implications.
Atom of Thoughts is an MCP server for structured AI reasoning. It decomposes problems into atomic thoughts (premise, reasoning, hypothesis, verification, conclusion), each tagged with a confidence score, forming a dependency chain. The server offers multiple tools (AoT-fast, AoT-full, atomcommands) to build and manage thought chains, and supports interactive visualization via a terminal TUI or browser (D3.js), allowing users to approve or reject atoms for human oversight. It is based on the paper 'Atom of Thoughts' (arXiv:2502.12018), written in TypeScript, and can be installed via npx, npm, Smithery, or Docker.
Tools
AoT-fast
Quick reasoning for tradeoffs, debugging, decisions, option evaluation (depth 3).
AoT-full
In-depth reasoning for plans, architecture, decomposition (depth 5).
atomcommands
Manage sessions, export graphs, poll approvals, and lifecycle.
Setup
Recommended npx: add {"command":"npx","args":["-y","@dioptx/mcp-atom-of-thoughts"]} to MCP config and restart your client. Also supports npm global, Smithery, and Docker.
claude_desktop_config.json
{
"mcpServers": {
"atom-of-thoughts": {
"command": "npx",
"args": [
"-y",
"@dioptx/mcp-atom-of-thoughts"
]
}
}
}
Fit and risk
Not for
- Simple Q&A without need for structured reasoning
- Scenarios where extra visualization or approval steps are unwelcome
- Users who don't need depth control or atom management
Required permissions
- Executes npx to fetch package, requires network
- Writes temporary files for visualization HTML
- May write to ~/Downloads for approval JSON
- Can start a local HTTP server for browser interaction
Risks and side effects
- Dependency on external npm packages, supply chain risk
- Visualization may open browser, privacy consideration
- Untrusted tool calls could produce unreliable reasoning
Troubleshooting
- Ensure Node.js version meets requirements (see package.json)
- Check MCP config for correct command and args
- If visualization doesn't open, check --viz setting and browser permissions
Use cases
Use AoT-fast for quick tech decisions (e.g., JWT vs session auth).
Use AoT-full for complex planning or architecture decomposition.
Visualize reasoning via TUI or browser and approve/reject atoms.
Supported clients
Claude CodeFull support
CursorFull support
WindsurfFull support