Skip to main content
OpenCode is the open-source, provider-agnostic terminal coding agent. The division of labor writes itself: OpenCode writes code · Nika captures the repeatable AI work as a checkable file. No plugin is required — OpenCode natively reads every surface nika init scaffolds.

​
Wire it (3 minutes)

1 · Teach the repo — OpenCode discovers AGENTS.md (project root, upward traversal, CLAUDE.md fallback) and repo skills under .agents/skills/ natively: 
nika init
That is the whole teaching step. The nika-authoring skill (author → check → repair loop) appears in OpenCode’s skill tool automatically. 2 · Wire the oracle — add the two-field stanza to opencode.json (project-local, or ~/.config/opencode/opencode.json globally): 
{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "nika": { "type": "local", "command": ["nika", "mcp"], "enabled": true }
  }
}
OpenCode’s own docs warn that heavy MCP servers blow the context budget — the Nika oracle is the lean counter-example: a handful of read-only validate/learn tools (what it exposes and refuses).

​
Prove the loop

Ask OpenCode:
Turn my weekly changelog digest into a Nika workflow from the chain template, check it, and run it with the mock model.
Expected tool trail: nika new --from chain → edit → nika check (static audit, nothing spent) → nika run … --model mock/echo (offline) → receipts in .nika/traces/.

​
When to reach for which

taskwho
implement/refactor/review codeOpenCode
a repeated AI pipeline (digest, triage, ETL, report, multi-model bench)a .nika.yaml OpenCode authors once
a one-off questionneither — just answer

​
Troubleshooting

  • Skill not listed — OpenCode discovers .agents/skills/*/SKILL.md from the git worktree root; re-check nika init ran at the repo root.
  • MCP server absent — validate opencode.json against its $schema; the nika binary must be on PATH (nika --version).
  • Which config wins — project-local opencode.json merges over the global file; keep the stanza in one place.