diff --git a/crates/miyabi-core/src/lib.rs b/crates/miyabi-core/src/lib.rs
index d94cf70..a5c060a 100644
--- a/crates/miyabi-core/src/lib.rs
+++ b/crates/miyabi-core/src/lib.rs
@@ -8,6 +8,7 @@ pub mod anthropic;
 pub mod tool;
 pub mod conversation;
 pub mod tools;
+pub mod token;
 
 pub use error::Error;
 pub use types::*;
@@ -23,3 +24,4 @@ pub use conversation::{
     Conversation, ConversationMessage, ConversationManager, ConversationMetadata, ConversationError,
 };
 pub use tools::{ReadTool, WriteTool, EditTool, create_file_tool_registry};
+pub use token::{TokenCounter, TokenUsage, ContextManager, ContextUsage, ModelLimits};
diff --git a/crates/miyabi-core/src/token.rs b/crates/miyabi-core/src/token.rs
new file mode 100644
index 0000000..7b634f6
--- /dev/null
+++ b/crates/miyabi-core/src/token.rs
@@ -0,0 +1,438 @@
+//! Token Counting and Context Window Management
+//!
+//! This module provides token estimation and context window management
+//! for Claude API conversations.
+
+use crate::anthropic::{ContentBlock, Message};
+use crate::conversation::{Conversation, ConversationMessage};
+use serde::{Deserialize, Serialize};
+
+/// Claude model context limits (in tokens)
+#[derive(Debug, Clone, Copy)]
+pub struct ModelLimits {
+    /// Maximum context window size
+    pub context_window: usize,
+    /// Maximum output tokens
+    pub max_output: usize,
+}
+
+impl ModelLimits {
+    /// Claude 3 Opus limits
+    pub const CLAUDE_3_OPUS: Self = Self {
+        context_window: 200_000,
+        max_output: 4_096,
+    };
+
+    /// Claude 3 Sonnet limits
+    pub const CLAUDE_3_SONNET: Self = Self {
+        context_window: 200_000,
+        max_output: 4_096,
+    };
+
+    /// Claude 3 Haiku limits
+    pub const CLAUDE_3_HAIKU: Self = Self {
+        context_window: 200_000,
+        max_output: 4_096,
+    };
+
+    /// Get limits for a model name
+    pub fn for_model(model: &str) -> Self {
+        if model.contains("opus") {
+            Self::CLAUDE_3_OPUS
+        } else if model.contains("haiku") {
+            Self::CLAUDE_3_HAIKU
+        } else {
+            Self::CLAUDE_3_SONNET
+        }
+    }
+}
+
+/// Token usage statistics
+#[derive(Debug, Clone, Default, Serialize, Deserialize)]
+pub struct TokenUsage {
+    /// Estimated input tokens
+    pub input_tokens: usize,
+    /// Output tokens (from API response)
+    pub output_tokens: usize,
+    /// Total tokens used
+    pub total_tokens: usize,
+}
+
+impl TokenUsage {
+    /// Create new usage
+    pub fn new(input: usize, output: usize) -> Self {
+        Self {
+            input_tokens: input,
+            output_tokens: output,
+            total_tokens: input + output,
+        }
+    }
+
+    /// Add usage
+    pub fn add(&mut self, other: &TokenUsage) {
+        self.input_tokens += other.input_tokens;
+        self.output_tokens += other.output_tokens;
+        self.total_tokens += other.total_tokens;
+    }
+}
+
+/// Token counter for estimating token usage
+#[derive(Debug, Clone)]
+pub struct TokenCounter {
+    /// Model limits
+    pub limits: ModelLimits,
+    /// Characters per token estimate (Claude uses ~4 chars/token on average)
+    pub chars_per_token: f32,
+}
+
+impl Default for TokenCounter {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl TokenCounter {
+    /// Create a new token counter with default settings
+    pub fn new() -> Self {
+        Self {
+            limits: ModelLimits::CLAUDE_3_SONNET,
+            chars_per_token: 4.0,
+        }
+    }
+
+    /// Create with specific model limits
+    pub fn with_model(model: &str) -> Self {
+        Self {
+            limits: ModelLimits::for_model(model),
+            chars_per_token: 4.0,
+        }
+    }
+
+    /// Estimate tokens for text
+    pub fn estimate_text(&self, text: &str) -> usize {
+        (text.len() as f32 / self.chars_per_token).ceil() as usize
+    }
+
+    /// Estimate tokens for a content block
+    pub fn estimate_content_block(&self, block: &ContentBlock) -> usize {
+        match block {
+            ContentBlock::Text { text } => self.estimate_text(text),
+            ContentBlock::ToolUse { name, input, .. } => {
+                let input_str = serde_json::to_string(input).unwrap_or_default();
+                self.estimate_text(name) + self.estimate_text(&input_str) + 20
+            }
+            ContentBlock::ToolResult { content, .. } => {
+                self.estimate_text(content) + 20
+            }
+        }
+    }
+
+    /// Estimate tokens for a message
+    pub fn estimate_message(&self, message: &Message) -> usize {
+        let content_tokens: usize = message
+            .content
+            .iter()
+            .map(|b| self.estimate_content_block(b))
+            .sum();
+
+        // Add overhead for message structure
+        content_tokens + 10
+    }
+
+    /// Estimate tokens for a conversation message
+    pub fn estimate_conversation_message(&self, message: &ConversationMessage) -> usize {
+        let content_tokens: usize = message
+            .content
+            .iter()
+            .map(|b| self.estimate_content_block(b))
+            .sum();
+
+        content_tokens + 10
+    }
+
+    /// Estimate tokens for a conversation
+    pub fn estimate_conversation(&self, conversation: &Conversation) -> usize {
+        let mut total = 0;
+
+        // System prompt
+        if let Some(ref prompt) = conversation.system_prompt {
+            total += self.estimate_text(prompt) + 10;
+        }
+
+        // Messages
+        for message in &conversation.messages {
+            total += self.estimate_conversation_message(message);
+        }
+
+        total
+    }
+
+    /// Get available tokens for output
+    pub fn available_output_tokens(&self, conversation: &Conversation) -> usize {
+        let input = self.estimate_conversation(conversation);
+        let available = self.limits.context_window.saturating_sub(input);
+        available.min(self.limits.max_output)
+    }
+
+    /// Check if conversation is within limits
+    pub fn within_limits(&self, conversation: &Conversation) -> bool {
+        let tokens = self.estimate_conversation(conversation);
+        tokens < self.limits.context_window - self.limits.max_output
+    }
+
+    /// Get context window size
+    pub fn context_window(&self) -> usize {
+        self.limits.context_window
+    }
+
+    /// Get max output tokens
+    pub fn max_output(&self) -> usize {
+        self.limits.max_output
+    }
+}
+
+/// Context window manager for automatic pruning
+#[derive(Debug, Clone)]
+pub struct ContextManager {
+    /// Token counter
+    pub counter: TokenCounter,
+    /// Target utilization (0.0 - 1.0)
+    pub target_utilization: f32,
+    /// Minimum messages to keep
+    pub min_messages: usize,
+}
+
+impl Default for ContextManager {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl ContextManager {
+    /// Create a new context manager
+    pub fn new() -> Self {
+        Self {
+            counter: TokenCounter::new(),
+            target_utilization: 0.8, // Keep 80% of context for input
+            min_messages: 2,         // Keep at least 2 messages
+        }
+    }
+
+    /// Create with specific model
+    pub fn with_model(model: &str) -> Self {
+        Self {
+            counter: TokenCounter::with_model(model),
+            ..Default::default()
+        }
+    }
+
+    /// Set target utilization
+    pub fn with_target_utilization(mut self, utilization: f32) -> Self {
+        self.target_utilization = utilization.clamp(0.1, 0.95);
+        self
+    }
+
+    /// Set minimum messages to keep
+    pub fn with_min_messages(mut self, min: usize) -> Self {
+        self.min_messages = min.max(1);
+        self
+    }
+
+    /// Get current token estimate
+    pub fn estimate_tokens(&self, conversation: &Conversation) -> usize {
+        self.counter.estimate_conversation(conversation)
+    }
+
+    /// Get target token limit
+    pub fn target_limit(&self) -> usize {
+        (self.counter.context_window() as f32 * self.target_utilization) as usize
+    }
+
+    /// Check if pruning is needed
+    pub fn needs_pruning(&self, conversation: &Conversation) -> bool {
+        self.estimate_tokens(conversation) > self.target_limit()
+    }
+
+    /// Prune conversation to fit within limits
+    ///
+    /// Removes oldest messages while keeping system prompt and minimum messages
+    pub fn prune(&self, conversation: &mut Conversation) -> usize {
+        let target = self.target_limit();
+        let mut removed = 0;
+
+        while self.estimate_tokens(conversation) > target
+            && conversation.message_count() > self.min_messages
+        {
+            conversation.messages.remove(0);
+            removed += 1;
+        }
+
+        removed
+    }
+
+    /// Get usage statistics
+    pub fn get_usage(&self, conversation: &Conversation) -> ContextUsage {
+        let tokens = self.estimate_tokens(conversation);
+        let limit = self.counter.context_window();
+        let available = self.counter.available_output_tokens(conversation);
+
+        ContextUsage {
+            used_tokens: tokens,
+            total_tokens: limit,
+            available_output: available,
+            utilization: tokens as f32 / limit as f32,
+            message_count: conversation.message_count(),
+        }
+    }
+}
+
+/// Context usage statistics
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ContextUsage {
+    /// Tokens used by current context
+    pub used_tokens: usize,
+    /// Total context window size
+    pub total_tokens: usize,
+    /// Available tokens for output
+    pub available_output: usize,
+    /// Utilization percentage (0.0 - 1.0)
+    pub utilization: f32,
+    /// Number of messages
+    pub message_count: usize,
+}
+
+impl ContextUsage {
+    /// Format as display string
+    pub fn display(&self) -> String {
+        format!(
+            "{}/{} tokens ({:.1}%) | {} messages",
+            self.used_tokens,
+            self.total_tokens,
+            self.utilization * 100.0,
+            self.message_count
+        )
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_token_estimation() {
+        let counter = TokenCounter::new();
+
+        // ~4 chars per token
+        let tokens = counter.estimate_text("Hello, World!"); // 13 chars
+        assert!(tokens >= 3 && tokens <= 5);
+    }
+
+    #[test]
+    fn test_model_limits() {
+        let limits = ModelLimits::for_model("claude-3-opus");
+        assert_eq!(limits.context_window, 200_000);
+
+        let limits = ModelLimits::for_model("claude-3-sonnet");
+        assert_eq!(limits.context_window, 200_000);
+    }
+
+    #[test]
+    fn test_conversation_estimation() {
+        let counter = TokenCounter::new();
+        let mut conv = Conversation::new()
+            .with_system_prompt("You are a helpful assistant");
+
+        conv.add_user_message("Hello");
+        conv.add_assistant_message("Hi there!");
+
+        let tokens = counter.estimate_conversation(&conv);
+        assert!(tokens > 0);
+    }
+
+    #[test]
+    fn test_within_limits() {
+        let counter = TokenCounter::new();
+        let conv = Conversation::new()
+            .with_system_prompt("Test");
+
+        assert!(counter.within_limits(&conv));
+    }
+
+    #[test]
+    fn test_context_manager_creation() {
+        let manager = ContextManager::new();
+        assert_eq!(manager.target_utilization, 0.8);
+    }
+
+    #[test]
+    fn test_context_manager_pruning() {
+        // Create manager with small model limits for testing
+        let mut manager = ContextManager::new();
+        manager.counter.limits = ModelLimits {
+            context_window: 100, // Very small for testing
+            max_output: 10,
+        };
+        manager.target_utilization = 0.5; // 50 token target
+
+        let mut conv = Conversation::new();
+        // Add messages that will exceed the limit
+        for i in 0..10 {
+            // Each message is roughly 30+ tokens
+            conv.add_user_message(format!("This is message number {} with substantial content that uses many tokens", i));
+        }
+
+        let initial = conv.message_count();
+        let removed = manager.prune(&mut conv);
+
+        assert!(removed > 0, "Expected messages to be removed, tokens: {}", manager.estimate_tokens(&conv));
+        assert!(conv.message_count() < initial);
+    }
+
+    #[test]
+    fn test_context_usage_display() {
+        let usage = ContextUsage {
+            used_tokens: 1000,
+            total_tokens: 200_000,
+            available_output: 4096,
+            utilization: 0.005,
+            message_count: 5,
+        };
+
+        let display = usage.display();
+        assert!(display.contains("1000"));
+        assert!(display.contains("200000"));
+    }
+
+    #[test]
+    fn test_available_output_tokens() {
+        let counter = TokenCounter::new();
+        let conv = Conversation::new();
+
+        let available = counter.available_output_tokens(&conv);
+        assert_eq!(available, counter.max_output());
+    }
+
+    #[test]
+    fn test_token_usage() {
+        let mut usage = TokenUsage::new(100, 50);
+        assert_eq!(usage.total_tokens, 150);
+
+        usage.add(&TokenUsage::new(50, 25));
+        assert_eq!(usage.total_tokens, 225);
+    }
+
+    #[test]
+    fn test_min_messages_preserved() {
+        let manager = ContextManager::new()
+            .with_target_utilization(0.001)
+            .with_min_messages(5);
+
+        let mut conv = Conversation::new();
+        for i in 0..10 {
+            conv.add_user_message(format!("Message {}", i));
+        }
+
+        manager.prune(&mut conv);
+        assert!(conv.message_count() >= 5);
+    }
+}