agentkit-mcp

MCP server that exposes the agent quality quartet — coderace, agentmd, agentlint, agentreflect — as tools inside Claude Code and any MCP-compatible client.

No context-switching. Run multi-agent code races, generate CLAUDE.md, lint diffs, and reflect on agent failures — all from within your AI coding session.

Prerequisites

Install the four Python CLIs:

pip install coderace agentmd-gen ai-agentlint ai-agentreflect

Or with pipx (recommended):

pipx install coderace
pipx install agentmd-gen
pipx install ai-agentlint
pipx install ai-agentreflect

Install

npm install -g agentkit-mcp

Or use directly without installing via npx agentkit-mcp.

Configure

Claude Code

claude mcp add agentkit-mcp npx agentkit-mcp

Manual .mcp.json

{
  "mcpServers": {
    "agentkit": {
      "command": "npx",
      "args": ["agentkit-mcp"]
    }
  }
}

Place this file in your project root or ~/.claude/.

Tools

coderace_race

Run a multi-agent coding race on a repository.

repo_path   (required)  Absolute path to the repository root
agents      (optional)  Comma-separated agent names: "codex,claude"
task        (optional)  Task description to pass to agents

Example prompt in Claude Code:

"Run a coderace race on /Users/me/myrepo with agents codex,claude to implement a binary search function."

---

coderace_review

Multi-lane AI code review on staged changes or a branch diff.

repo_path      (required)  Absolute path to the repository root
diff_source    (optional)  staged | branch | diff
lanes          (optional)  Comma-separated lane names
cross_reviewers (optional) Number of cross-reviewers

Example:

"Review the staged changes in /Users/me/myrepo using coderace_review."

---

agentmd_generate

Generate a CLAUDE.md / AGENTS.md from a repository. Inspects code structure, dependencies, and conventions.

repo_path  (required)  Absolute path to the repository root
output     (optional)  Output file path (default: CLAUDE.md in repo root)
minimal    (optional)  Generate a minimal file
tiered     (optional)  Generate a tiered/layered file

Example:

"Generate a CLAUDE.md for /Users/me/myrepo."

---

agentmd_diff

Show a diff between the current CLAUDE.md and a freshly generated one. Spot what's gone stale.

repo_path  (required)  Absolute path to the repository root
output     (optional)  Output file path

---

agentlint_check

Lint a git diff for agent-safety issues: missing context, ambiguous instructions, prompt-injection risks.

diff  (required)  Git diff content (paste text) OR path to a .diff file

Example:

"Lint this diff with agentlint_check: "

---

agentlint_check_context

Lint an AGENTS.md or CLAUDE.md for freshness, completeness, and structural quality. Returns a freshness score and actionable suggestions.

file  (required)  Path to AGENTS.md or CLAUDE.md

Example:

"Check if my AGENTS.md at /Users/me/myrepo/AGENTS.md is up to date."

---

agentreflect_generate

Generate agent improvement suggestions from notes, pytest output, or git history. Optionally apply them to AGENTS.md / CLAUDE.md.

from_notes   (optional)  Free-form notes to reflect on
from_pytest  (optional)  Path to pytest output file
from_git     (optional)  Reflect from recent git log
apply        (optional)  Path to apply suggestions to
yes          (optional)  Auto-accept all suggestions

Example:

"Use agentreflect_generate with these notes: 'Tests failed 3 times, agent missed the cwd convention.'"

---

Development

git clone https://github.com/mikiships/agentkit-mcp
cd agentkit-mcp
npm install
npm run build
npm test

To test locally as an MCP server:

node dist/index.js

License

MIT