Skip to main content
Claude Code gets two Nika surfaces: the MCP oracle — schema, validation and cost answers without shell round-trips (what it exposes and refuses) — and the first-party plugin. The binary comes first: brew install supernovae-st/tap/nika (other paths).

​
What it writes

~/.claude.json — the nika entry lands under mcpServers; every unrelated server and setting survives, and re-running prints · current (idempotent). Prefer manual? Add the same block yourself:

​
The plugin (skill + commands)

Then nika init in the repo — the agent-teaching surface (AGENTS.md, rules, the repo skill) is what makes the oracle useful (agents guide).

​
Keep it fresh

Two things move, on the same release train — the plugin never updates the binary, and brew never updates the plugin. Claude Code’s plugin update is TWO rungs (proven live: an install once sat three releases behind, silently, with a fresh clone right next to it):
nika doctor answers “is my suite coherent?” — from v0.107 it reads the install rung (what your sessions actually load), names any kit lagging or leading the binary’s train, and prints both rungs when Claude Code needs the climb. Drift is advisory (warn · exit 0), so it never breaks automation.

​
Verify

claude mcp list shows nika, and the tool list carries the nine nika_* tools. Then prove the loop end-to-end: integrate in 10 minutes.