kimi-for-claude

claude is a genius at managing, but using it for raw coding is a token drain. kimi has claude-level quality for a fraction of the price, but orchestrating both manually is a headache

so i built a bridge

claude acts as the supervisor, delegating the heavy coding to kimi in the background. you get claude's brain guiding the project, with kimi's budget-friendly execution

What it does

Send Claude a task too large, too tedious, or better run in parallel, Claude hands it off to Kimi with full context, waits for the result (or doesn't), then picks up the conversation. Kimi gets his own workspace, his own tools (shell, file ops, web search, all skills), and runs autonomously.

Tools

Every tool supports:

• work_dir — Kimi's workspace. If it's a git repo, results include a one-line summary of files he changed.
• readonly — Plan mode: Kimi investigates and reports, no writes.
• background — Returns a job ID instantly so Claude can keep working in parallel.

Results are truncated at 6k chars by design, Claude asks Kimi to summarize rather than reading full output, keeping token usage lean.

Requirements

• Claude Code
• Kimi CLI installed and authenticated (kimi login) and of course a funded/subscribed account

Setup

Pick whichever runtime you have.

Node.js (18+) — from npm (easiest)

claude mcp add --scope user kimi -- npx -y kimi-for-claude

Or install globally first:

npm install -g kimi-for-claude
claude mcp add --scope user kimi -- kimi-for-claude

Node.js (18+) — from source

1. Clone the repo

git clone https://github.com/7D-codes/kimi-for-claude
cd kimi-for-claude
npm install

2. Register with Claude Code

claude mcp add --scope user kimi -- node "$PWD/server.js"

Python (3.10+) + uv

1. Clone the repo

git clone https://github.com/7D-codes/kimi-for-claude
cd kimi-for-claude

2. Register with Claude Code

claude mcp add --scope user kimi -- uv run --script "$PWD/server.py"

3. Verify it connected

claude mcp list
# kimi: ... - ✔ Connected

4. Start a new Claude Code session — the tools won't appear in your current session, only new ones.

That's it. kimi_delegate, kimi_continue, kimi_status, and kimi_cancel are now in Claude's toolbox.

Usage examples

Delegate a task

Have Kimi write all the unit tests for this module while you review the architecture.

Research in parallel

Ask Kimi to find the best approach for rate-limiting in Redis — summarize when done.

Iterate

Claude delegates → reviews Kimi's output → kimi_continue to request fixes → done.

Background job

Claude fires off a long build task to Kimi with background=true, keeps working, checks kimi_status when convenient.

How it works

server.js and server.py are both single-file MCP stdio servers. server.py uses FastMCP with inline dependencies (PEP 723), run by uv — no install step. server.js uses the official @modelcontextprotocol/sdk with npm install. Both shell out to:

kimi --prompt "<task>" [--continue]      # run in cwd=work_dir

Non-interactive --prompt mode auto-executes Kimi's tool calls (no --yolo needed; --yolo/--plan are rejected when combined with --prompt). The working directory is set via the subprocess cwd (the old -w flag was removed). readonly mode is enforced by a prompt-level instruction since there is no non-interactive plan flag. The safety gate is Claude Code's permission prompt on each MCP tool call — you see every task before it reaches Kimi.

Remove

claude mcp remove kimi -s user