@kawacode/mcp

kawa.mcp

Reference implementation of the Kawa Code MCP protocol.

Overview

The Kawa Code MCP server provides the communication layer used by Kawa Code to record and align development intent between developers and AI systems.

It enables:

• Persistent AI reasoning context • Intent tracking during development workflows • Alignment between human and AI decisions over time

This repository contains the reference implementation of the Kawa MCP server used by Kawa Code tools.

This MCP server enables AI coding assistants to understand what you're working on and maintain context across sessions. It connects to the Kawa Code desktop application to provide:

• Intent tracking: Create and manage development intents with decision history
• Team collaboration: See what teammates are working on, detect conflicts
• Decision recording: Track architectural decisions and trade-offs with constraint validation
• Code block assignment: Associate code changes with intents for better commit history

Prerequisites

Required

• Node.js >= 18.0.0 — runtime for the MCP server
• Kawa Code desktop app running — kawa.mcp is a thin MCP-to-IPC adapter; all git operations, storage, and API communication happen in Kawa Code

Optional (for history inference)

• Anthropic API key — your own Claude API key, passed as a parameter to the inference tools
• GitHub CLI (gh) — enables richer data tiers (PR descriptions, review comments, issue discussions). Without gh, tiers 2 and 4 are skipped automatically

Installation

Add the MCP server to Claude Code (available across all your projects):

claude mcp add kawa-intents --scope user -- npx -y @kawacode/mcp

Or for a single project, create a .mcp.json file in your project root (recommended for teams — commit it to git):

{
  "mcpServers": {
    "kawa-intents": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@kawacode/mcp"]
    }
  }
}

Cursor AI setup

Add to your Cursor MCP configuration (~/.cursor/mcp.json):

{
  "mcpServers": {
    "kawa-intents": {
      "command": "npx",
      "args": ["-y", "@kawacode/mcp"]
    }
  }
}

Quick Start

• Start Kawa Code: Launch the Kawa Code desktop app and log in
• Install MCP: Run claude mcp add kawa-intents --scope user -- npx -y @kawacode/mcp
• Restart AI: Restart Claude Code or Cursor to load the MCP server
• Test connection: The server will try to connect to Kawa Code on startup
• Start coding: You'll start seeing check_active_intent at the start of your work with Claude Code

Setting Up CLAUDE.md

Quick prompt to tell Claude Code to set up everything:

Read the CLAUDE.md.example file from the @kawacode/mcp package and create a CLAUDE.md in this project's root. Fill in the repoOrigin and repoPath with the actual values from this repository's git config. Fill in the Project Overview with a brief description of this project.

After this, you can start working using your usual workflow. Kawa Code will automatically start improving your LLM's code generation quality.

To benefit from Kawa Code's intent-driven development immediately, tell Claude Code to infer code decisions from your git history:

Please run infer_history with max 3000 commits.

Tier 5 (the default) extracts the most context. Tiers are cumulative — each includes all data from lower tiers. Tiers 2 and 4 add PR/issue context from GitHub/GitLab and require the gh or glab CLI; they're automatically skipped if unavailable.

This can take a while, depending on how many commits you asked it to analyze. You'll see a progress bar in Kawa Code application. Once this is done, you can open the project in Kawa Code (or you can use one of Kawa Code extensions for Visual Studio Code, emacs, or vim) to see the intents at every step in the code evolution. Kawa Code will now select the relevant intents and micro-decisions made to speed up code generation, issue troubleshooting, bug fixing etc. This relevant context will also help generating a more correct solution to your prompt request.

Usage

The MCP server works together with Kawa Code, AI code generators such as Cursor, Claude Code, and the Kawa Code extensions.

Key Features

• Context Persistence: Never lose track of what you were working on across AI sessions
• Smart Context Retrieval: Relevance-based context loading - only fetch what's needed for the current task
• Zero-Knowledge Encryption: Code blocks encrypted client-side before cloud sync, API cannot decrypt
• Team Conflict Detection: Know when teammates are working on the same files/lines
• Decision Tracking: Record architectural decisions with constraint validation and conflict detection
• Commit Integration: Link all code changes to intent context for better git history
• Cross-Platform: Works with Claude Code and Cursor AI via MCP protocol

Available Tools

Context & Discovery

Tool	Description
get_relevant_context	Get context relevant to a specific task

Intent Management

Tool	Description
check_active_intent	Check if there's an active intent before starting new work
create_and_activate_intent	Create and activate a new intent for a development task
get_intents_for_file	Get intents (team and self) affecting a specific file
get_intents_for_lines	Get intents affecting specific line ranges in a file
assign_blocks_to_intent	Assign modified line ranges to the active intent
get_intent_changes	Get uncommitted changes for the active intent
complete_intent	Complete an intent (committed/done/abandoned)
list_team_intents	List what teammates are currently working on

Decision Recording

Tool	Description
record_decision	Record an architectural decision with rationale and constraint validation
get_session_decisions	Get decisions recorded during the current session
get_project_decisions	Get all decisions across all intents for the project
edit_session_decision	Edit or delete a decision before intent completion
detect_intent_conflicts	Detect if current intent decisions conflict with team decisions

History Inference

Tool	Description
infer_history	Analyze git commit history to extract development stories and decisions
evolve_decisions	Build a decision evolution graph from previously extracted stories

Lightweight Logging

Tool	Description
log_work	Log completed work without the full intent lifecycle — use for quick fixes and trivial changes

MCP Capabilities

Prompts

The server exposes prompts that can be loaded into your AI coding session:

• implementation_workflow: Standard workflow for implementing code changes with intent tracking. Provides step-by-step guidance on checking for active intents, creating new intents, checking for conflicts, and assigning blocks.

Resources

The server exposes resources that can be monitored:

• kawa://intent/active: Real-time view of the currently active intent for the connected repository (JSON format)

History Inference

Two MCP tools analyze git commit history to extract structured development knowledge — useful for bootstrapping a repository with historical context.

infer_history

Runs the full pipeline automatically: infer → evolve → persist.

• Pass 1: Groups commits into coherent development stories with value hints (high/low/none)
• Pass 2: Deep analysis of high/low-value stories to extract architectural decisions and lessons learned
• Evolution: Curates decisions by finding relationships (supersedes, reinforces, contradicts, specializes)
• Persist: Stores curated stories as intents with decisions and lessons (auto-syncs to cloud)

The pipeline runs asynchronously inside Kawa Code. Progress is shown in the Kawa Code desktop app via a progress bar. The pipeline supports checkpointing — if interrupted, re-running resumes from where it left off.

Usage in Claude Code:

Use the infer_history tool with estimateOnly: true to preview the cost first,
then run it with estimateOnly: false.

Parameters:

Parameter	Type	Default	Description
repoPath	string	(required)	Local path to the repository root
apiKey	string	(required)	Your Anthropic API key
commits	number	50	Number of recent commits to analyze
tier	number	4	Data enrichment tier (1-5, see below)
model	string	claude-sonnet-4-20250514	Anthropic model to use
maxStories	number	0	Limit stories to analyze in Pass 2 (0 = unlimited)
allowCommitSplitting	boolean	false	Allow splitting a commit into multiple stories when it contains unrelated changes (recommended for repos with messy commit history)
contextIssues	boolean	false	Include context issues from commit date range (tier 4 only)
estimateOnly	boolean	false	Preview token cost without running the pipeline

evolve_decisions

Builds a decision evolution graph from previously extracted stories — identifying how decisions relate across stories over time. Note: infer_history already chains evolve + persist automatically. Use this tool only if you want to run evolution separately on a pre-existing set of stories.

• Bucketing: Groups stories by file overlap and keyword similarity
• Edge classification: Uses LLM to identify relationships (supersedes, reinforces, contradicts, specializes)
• Annotation: Labels each decision as stable, orphan, evolved, or abandoned
• Curation: Keeps stable + orphan decisions, drops evolved + abandoned

If repoPath is provided, curated stories are automatically persisted as intents with decisions and lessons after evolution completes.

Parameters:

Parameter	Type	Default	Description
stories	array	(required)	Story objects from a previous infer_history run
apiKey	string	(required)	Your Anthropic API key
model	string	claude-haiku-4-5-20251001	Anthropic model (cheaper model recommended)
repoPath	string	(optional)	Local path to repository root (required for auto-persist)
repoOrigin	string	(optional)	Git remote origin URL (auto-detected from repoPath if not provided)

Data Tiers

Each tier adds more context for better inference. Higher tiers require the gh CLI authenticated.

Tier	Data source	Requires gh
1	Commit messages + numstat	No
2	+ PR descriptions and review comments	Yes
3	+ Diffs for revert commits	No
4	+ Referenced GitHub issues (default)	Yes
5	+ Diffs for all commits with annotation extraction	No

Development

# Watch mode (auto-rebuild on file changes)
npm run dev

# Build TypeScript to JavaScript
npm run build

# Clean build artifacts
npm run clean

# Run the MCP server directly
npm start

Testing the MCP Server

To test the MCP server without integrating it into an AI assistant:

• Build the project: npm run build
• Run the server: npm start
• The server communicates via stdio (standard input/output)
• You can send MCP protocol messages via stdin to test tool functionality

Development Tips

• Use npm run dev to auto-rebuild during development
• Check stderr for server logs (stdout is reserved for MCP protocol)
• Ensure Kawa Code is running before testing

Architecture

Claude Code / Cursor AI
    ↓ MCP Protocol (stdio)
kawa.mcp (this server)
    ↓ Huginn IPC (Unix socket / Named pipe)
Kawa Code Desktop App
    ├─ Gardener Module (Rust)
    │   └─ Intent/Decision storage
    └─ HTTP Client
        ↓ REST + SSE
    Kawa API (cloud)
        └─ Team sync & encryption

The MCP server communicates with Kawa Code using the Huginn IPC protocol:

• Context queries: Intents, decisions, relevant context
• Intent operations: Create, update, assign blocks, complete
• Decision tracking: Record, retrieve, edit, conflict detection

Kawa Code's Gardener module handles all git operations, diff generation, and local storage of encrypted data.

Contributing

Contributions are welcome. Please read CONTRIBUTING.md and CLA.md.

License

This project is source-available under the Kawa Code Source Available License.

You may run and modify the software for personal or internal use.

See LICENSE for details.