Kawa Code MCP

What Kawa Code MCP does

Kawa Code MCP MCP server gives AI coding assistants access to team-aware repository context through the Kawa Code desktop application. It helps an assistant continue work across sessions, branches, and machines without reconstructing the project background each time. It also stores architectural decisions together with their reasoning, so later sessions can review the choices behind the current code.

The coordination features cover active work as well as historical context. Agents and teammates can see overlapping in-progress edits in working copies before either change is committed. Commits are associated with the intent that motivated them, adding rationale alongside the changes shown by Git.

How it works

The package is an MCP-to-IPC adapter. Kawa Code performs the Git operations, data storage, and API communication, while the MCP process exposes that functionality to an AI assistant. The setup therefore depends on the Kawa Code desktop application being open for the project.

Context retrieval is relevance-based rather than loading every recorded item. The documented get_relevant_context capability can surface information related to the current task. Intent tracking is session-specific, allowing several agents to work on one repository with separate active focuses.

For parallel agent work, each worker should use its own Git worktree. Separate worktrees prevent agents from overwriting one another without Git being able to represent the conflict. Kawa's coordination features then report overlapping edits across those working copies. The live edit-level signal is provided by the Stop hook; the former pre-edit decision-check hook was retired on August 14, 2026.

Setup and configuration

Kawa Code MCP MCP server requires Node.js 18 or later and a running Kawa Code desktop application. A project-level .mcp.json can run it with npx -y @kawacode/mcp, which is suitable for committing shared configuration to a repository. Claude Code can add the same package with its MCP command. Cursor can use a global installation and a kawacode-mcp command in ~/.cursor/mcp.json.

A global npm installation does not update itself. Upgrade it with npm update -g @kawacode/mcp. An Anthropic API key is optional and is passed to history-inference tools as a parameter. Installing GitHub CLI is also optional; it enables richer history sources such as pull request descriptions, review comments, and issue discussions. Without it, the corresponding data tiers are skipped.

Tools and capabilities

• Maintains intent and context across sessions, branches, and machines.
• Records architectural decisions, trade-offs, and abandoned approaches with their reasoning.
• Retrieves task-relevant context instead of requiring all history to be loaded.
• Detects teammates editing overlapping files or lines before commits create merge conflicts.
• Associates commits with the intent behind them.
• Supports safe-tier overlap resolution when operating in an agent-owned worktree.
• Encrypts code blocks client-side before synchronization; Kawa cloud cannot decrypt that code.

Limitations and notes

Kawa Code MCP MCP server does not replace the Kawa Code desktop application: the application must be running, and it owns storage, Git operations, and API communication. Agents sharing one checkout can still overwrite one another without conflict markers, so worktrees are required for isolated parallel agent sessions.

The retired pre-edit hook should not be newly configured. Existing installations continue to work until removed manually, but current edit coordination uses the Stop hook. Compiled-language build output should not be shared between worktrees through symlinks because separate checkouts can overwrite the same artifacts. GitHub CLI and the Anthropic API key extend optional history features but are not required for the base MCP process.