Kimi Code CLI reads MCP servers
from mcp.json at two levels: user ~/.kimi-code/mcp.json (shared across
projects) and project .kimi-code/mcp.json (a same-named project entry
wins). The binary comes first: brew install supernovae-st/tap/nika
(other paths).
Add the nika entry to ~/.kimi-code/mcp.json:
Or conversationally: /mcp-config in the TUI adds servers without
hand-editing JSON · /mcp shows connection status.
The legacy kimi-cli taught kimi mcp add … — that subcommand does not
exist in kimi-code, its successor. The file above is the contract.
Kimi’s own docs also warn that project-level stdio entries execute when a
session starts: trust-scope the project file accordingly.
Prove it (deterministic · verified live 2026-07-31 · kimi-code 0.30.0 · nika 0.106.1)
A headless run emits one JSON object per line, tool calls included — the
oracle answering IS the receipt:
The stream carries
"tool_calls":[{"function":{"name":"mcp__nika__nika_canon"}}] — the
server loaded and the agent reached it. kimi doctor validates the
config files (exit 0/1).
The repo side
nika init writes the .agents/skills authoring skill — kimi-code scans
four skill tiers including project and user .agents/skills/, so the
skill surfaces as /skill:nika-authoring with zero extra wiring. Note
that kimi-code does not auto-load AGENTS.md today: the skill and the
oracle carry the teaching (agents guide).
Division of labor unchanged: Kimi orchestrates and writes code · Nika
runs the repeatable work as a checkable file with receipts — nika check
before, --max-cost-usd during, nika trace verify after. 
