MCP server that exposes a "think" tool (plus optional criticize / plan / search) enabling explicit intermediate reasoning steps for agentic LLMs.
https://github.com/Rai220/think-mcpYour AI agents are making mistakes they shouldn't. They're jumping to conclusions, missing obvious policy violations, and fumbling multi-step workflows because they lack one critical capability: the ability to pause and think.
The think-mcp server fixes this by implementing Anthropic's "think" tool research as a drop-in MCP server. Your agents can now record explicit reasoning steps, backtrack when needed, and work through complex decisions methodically.
You've probably seen this: your agent gets tool results, immediately acts on them without proper analysis, then fails spectacularly. Or it violates clear policies because it never stopped to check compliance. These failures happen because most LLMs lack structured reasoning capabilities during tool use.
Anthropic's research showed that adding a simple "think" tool improves evaluation metrics across the board - even for models without native reasoning skills.
Core "think" tool: Your agent can pause, record thoughts, and reason through complex situations without affecting the environment. It's like giving your agent a notepad for working through problems.
Advanced reasoning suite (optional): Add criticize, plan, and search tools for comprehensive problem-solving workflows.
Zero friction integration: One config block and your agents instantly gain structured reasoning capabilities.
Policy-heavy environments: Before executing financial transactions, your agent thinks through compliance requirements: "Let me verify this transfer meets the daily limit policy and has proper authorization..."
Multi-step data analysis: When processing API results, your agent reasons: "The user count dropped 15% - let me think about potential causes before suggesting solutions..."
Complex tool chains: Your agent plans: "I need to fetch user data, validate permissions, then update records - let me think through the proper sequence..."
Add to your MCP server configuration:
"mcpServers": {
"think-mcp": {
"command": "uvx",
"args": ["think-mcp"],
"enabled": true
}
}
For the full reasoning suite with search capabilities:
"mcpServers": {
"think-mcp": {
"command": "uvx",
"args": ["think-mcp", "--advanced"],
"enabled": true,
"env": {
"TAVILY_API_KEY": "your-key-here"
}
}
}
Your agents immediately gain access to structured reasoning tools. No code changes, no complex setup.
The think tool doesn't change your environment or data - it simply logs reasoning steps. This creates an explicit thought process that prevents rushed decisions and enables better multi-step planning. Your agents become more reliable without sacrificing speed.
Think-mcp transforms impulsive AI agents into methodical problem-solvers. Install it once, and every complex workflow becomes more reliable.