Kage

What kage-core/Kage MCP server does

The kage-core/Kage MCP server manages repository memory for coding agents and can coordinate delegated implementation work. It captures decisions, bug fixes, conventions, and codebase knowledge as files under .agent_memory/. These files use the Open Knowledge Format (OKF), remain readable as plain Markdown, and can be reviewed and shared through Git.

Memory entries include citations to the code they describe. Kage checks those citations when memory is written and recalled, and can withhold knowledge when the referenced code has changed. It also performs diff-time stale checks, helping prevent an agent from reusing a claim that no longer matches the repository.

How it works

After installation, Kage creates the memory directory, builds a code graph, writes agent guidance in AGENTS.md or CLAUDE.md, detects supported agents, and configures Git-related files including the packet merge driver. Retrieval runs locally using BM25 and vectors, with no account, hosted database, or API key required.

The kage-core/Kage MCP server can brief an agent from the repository's memory and run the work in a separate Git worktree. A run may be a single delegated task or part of a multi-wave goal. Kage then runs the checks used for its receipt itself rather than relying on the agent's report. The documented checks include tests, changed-line limits, and validation that formally cited paths exist in the worktree. A merge command lands a run only after its claim meets the configured checks and records what the agent learned.

Memory is committed with the repository, so knowledge discovered by one agent or developer can be used by later sessions and teammates. The repository also provides a local room, run board, and memory view through the daemon and app commands.

Setup and configuration

Run the installer from the repository root:

npx -y @kage-core/kage-graph-mcp install

Node.js 18 or newer is required. Restart the coding agent after installation. The installer handles initial memory setup, code-graph construction, agent wiring, policy files, Git ignore configuration, and the packet merge driver.

For narrower setups, the README documents commands for initializing memory without agent wiring, configuring one supported agent, and verifying an agent harness. The project also documents an app mode and a macOS arm64 desktop shell. LAN access is available only after manually setting "lan": true in .agent_memory/config.json; the README states that there is currently no --lan flag or app toggle.

Tools and capabilities

The kage-core/Kage MCP server supports workflows represented by these CLI operations:

• kage room for an interactive project room that briefs and hires agents.
• kage dispatch for sending an intent to a selected agent.
• kage runs and kage review for tracking and inspecting delegated work.
• kage merge for accepting a verified run and ratifying its learned memory.
• kage init, kage setup, and kage setup verify-agent for initialization and agent configuration.
• kage app for the local daemon-backed room, run board, and memory view.

The repository lists compatibility with Claude Desktop, Cursor, Windsurf, Cline, and other coding-agent clients. It does not require a paid API key.

Limitations and notes

The desktop packaging described in the README is macOS arm64 only; Windows and Linux packages are not built yet. Unsigned macOS builds can show an unidentified-developer prompt on first launch. LAN mode requires manual configuration and a pairing secret on requests. Kage's memory and retrieval are local, while shared use depends on committing the memory files to Git.