Mcts Reasoning — AI skill for Claude Code

MCTS-Reasoning

Monte Carlo Tree Search for LLM reasoning. Explores multiple reasoning paths using UCB1 to balance exploration and exploitation, with deterministic verifiers providing the reward signal.

See the [blog post](https://metafunctor.com/post/2026-04-01-i-spent-048-to-find-out-when-mcts-actually-works-for-llm-reasoning/) for experimental results and analysis.

When This Helps

MCTS adds value over best-of-N sampling under three conditions:

1. A deterministic verifier scores solutions (not a learned reward model).
2. The verifier provides partial credit (a gradient, not just pass/fail).
3. The problem is hard enough that blind sampling can't reliably solve it.

On constraint satisfaction problems with 6-8 variables, MCTS(5+3) solved 100% of problems vs 90% for best-of-N and self-consistency, both using the same total budget of 8 solution attempts. Total API cost: $0.48.

Install

pip install -e .                    # Core
pip install -e ".[anthropic]"       # + Anthropic provider
pip install -e ".[openai]"          # + OpenAI provider
pip install -e ".[all]"             # Everything

Quick Start

CLI

mcts-reason "What is 15*7+23?" --answer 128 --simulations 10
mcts-reason "Question" --provider anthropic --model claude-haiku-4-5-20251001

Python

from mcts_reasoning.mcts import MCTS
from mcts_reasoning.generator import LLMGenerator
from mcts_reasoning.evaluator import NumericEvaluator
from mcts_reasoning.providers.anthropic import AnthropicProvider

llm = AnthropicProvider()
gen = LLMGenerator(llm=llm, temperature=0.7, max_tokens=150)
ev = NumericEvaluator(ground_truth=128)

mcts = MCTS(generator=gen, evaluator=ev, max_rollout_depth=5)
state = mcts.search("What is 15*7+23?", simulations=10)

Run Experiments

pip install -e ".[anthropic]"
export ANTHROPIC_API_KEY=your-key
python experiments/run_csp.py --hard --budget 1.00

The Anthropic provider tracks token usage and enfor