kage — Claude Code MCP server: agent memory that gets checked
Keeps a repo's decisions, bug findings and deploy notes as Markdown, then verifies them against the code so agents stop reusing what went stale.
claude mcp add kage -- npx -y @kage-core/kage-graph-mcpWhat kage does
Kage builds project memory out of what agents do while they work: why a piece of code looks the way it does, how a nasty bug was tracked down, what to remember before a deploy. Entries are plain Markdown under .agent_memory/ and travel with the repo through git, so teammates read the same set. The format follows Google Open Knowledge Format, which keeps you free to move elsewhere.
The orchestrator takes an intent, briefs a coding agent from that memory and runs it inside its own git worktree. Tests, diff size and the existence of every cited path are commands Kage runs on its own; the agent's own summary carries no weight. kage merge lets the code through only after the claim holds up, and records what the agent picked up along the way.
Setup is one command from inside the repo: npx -y @kage-core/kage-graph-mcp install. It creates .agent_memory/, builds the code graph, writes the policy into AGENTS.md or CLAUDE.md, wires the agents it finds and configures .gitignore plus the merge driver. Node.js 18+ is required. No account, no API key. Restart your agent afterwards, otherwise the new policy never loads.
What it helps with
- keep a hard-won bug analysis from vanishing after the merge
- stop stale notes from reaching the agent as current truth
- confirm the tests actually ran instead of trusting a written claim
- hand a new teammate deploy context through git
- run a delegated task in an isolated worktree and read the receipt
Example prompts for the agent
- Set up Kage in this repo: run
npx -y @kage-core/kage-graph-mcp install, then tell me to restart you. - Record why we dropped the per-request cache and check that note against the code.
- Run
kage dispatch "fix the failing auth test" --agent claudeand show me the receipt.
How to install kage
Platform CLI
claude mcp add kage -- npx -y @kage-core/kage-graph-mcpPlatform CLI
codex mcp add kage -- npx -y @kage-core/kage-graph-mcpConfig file
{
"mcpServers": {
"kage": {
"command": "npx",
"args": [
"-y",
"@kage-core/kage-graph-mcp"
]
}
}
}Config file
{
"servers": {
"kage": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"@kage-core/kage-graph-mcp"
]
}
}
}Platform CLI
gemini mcp add kage npx -y @kage-core/kage-graph-mcpStep by step, with commands for each platform: How to add an MCP server to Claude Code, Codex, Cursor, VS Code, Gemini CLI and more
Questions about kage
How do I install kage in Claude Code?
Run claude mcp add kage -- npx -y @kage-core/kage-graph-mcp, then restart the agent. You can also run npx -y @kage-core/kage-graph-mcp install inside the repo, which writes the policy and wires the agents it detects.
What does kage do?
It keeps project memory as Markdown files in your repo and checks each entry against the code, so outdated notes never make it into an agent brief. It also runs delegated tasks in a separate git worktree and executes the checks itself.
Do I need an account or an API key?
No. Node.js 18+ and a git repo are enough. Memory lives in files and is shared through git, with no external service involved.
