Skip to content

Use mokata with other AI agents

mokata is not Claude-Code-only. It runs behind a thin harness boundary, so the same engine — the spec-driven pipeline, the completeness gate, the knowledge graph, self-healing memory, governance — works under the AI coding agent you already use. mokata wires its command set into each agent's native command surface (as /<name> — e.g. /brainstorm, /spec) and degrades clearly where an agent lacks a capability (it never pretends).

One command per agent: mokata setup <agent> (human-gated, idempotent, reversible with mokata unsetup <agent>). See the live matrix any time with mokata harness.

What's wired, what degrades (per agent)

Agent Commands → native surface Context PreToolUse hooks Subagents MCP server
Claude Code (claude) ✅ .claude/commands/*.md ✅ ✅ secret-guard + run-state gate-guard + SessionStart briefing ✅ native fan-out ✅ auto (.mcp.json)
Codex (codex) ✅ .codex/prompts/*.md ✅ ❌ degrades ❌ degrades manual (TOML config)
Cursor (cursor) ✅ .cursor/commands/*.md ✅ .cursor/rules ❌ degrades ❌ degrades ✅ auto (.cursor/mcp.json)
GitHub Copilot (copilot) ✅ .github/prompts/*.prompt.md ✅ copilot-instructions.md ❌ degrades ❌ degrades manual (VS Code mcp.json)
Windsurf (windsurf) ✅ .windsurf/workflows/*.md ✅ .windsurf/rules ❌ degrades ❌ degrades manual (~/.codeium/windsurf/)
Gemini CLI (gemini) ✅ .gemini/commands/*.toml ✅ GEMINI.md ❌ degrades ❌ degrades ✅ auto (.gemini/settings.json)
Aider (aider) ❌ reference prompts only¹ ✅ conventions / --read ❌ degrades ❌ degrades manual / none

¹ Aider has no user-authored slash-command file system (its /commands are built-in), so mokata declares commands absent for it and ships the mokata prompts as reference files (.aider/mokata-commands/) you /read or paste — never pretended to be native commands.

What "degrades clearly" means

mokata declares a capability absent unless it can verify the agent really supports it (the Stage-52 inviolable: when unsure, declare absent). When the engine needs an absent capability the harness boundary returns a clear result that names the missing capability — never a silent no-op of a gate:

  • No PreToolUse hooks (every agent except Claude Code): only Claude Code declares the hooks capability, so on Codex, Cursor, Copilot, Windsurf, Gemini CLI, and Aider neither PreToolUse hook is wired. Two distinct consequences, and they are not equally recoverable:

    • The secret-guard doesn't run on native writes — but durable-write protection still holds where it counts: mokata's own gated CLI/MCP WriteGate scans for secrets, human-gates, and audits every durable write regardless of the hook. Route writes through mokata (or its MCP tools), not raw, and you keep this.
    • The run-state gates enforce NOTHING. All four — approach-approval, spec-persisted, no-code-without-failing-test, and spec-scope — live only in the gate-guard hook. With no hook there is no interception of a native Write/Edit, so on these agents the agent can write implementation code before an approach is approved, before the spec is emitted, before a failing test is on record, or outside the spec's authorized surface — and nothing will stop it. The mokata commands still run the pipeline, and /review still catches divergence after the fact, but the hard seatbelt is Claude-Code-only. Nothing mokata ships works around this; there is no non-hook substitute. If enforced TDD is why you're here, use Claude Code.

  • No native subagents: parallel fan-out falls back to mokata's sequential gated flow (the cost estimate + two-stage review still apply); nothing is skipped.

  • MCP is a manual step: register the mokata-mcp server with the agent per its docs (mokata states this in the setup plan rather than half-wiring a schema it can't be sure of).

Quickstart

# wire mokata into your agent (shows exactly what it will write, then asks)
mokata setup cursor          # or: copilot · windsurf · gemini · aider · codex · claude
# … point the agent at the generated commands; register mokata-mcp if the table says "manual"
mokata harness               # see the full capability matrix any time
mokata unsetup cursor        # reverse it cleanly — no residue

Setup is human-gated (it previews every file it will write/merge, then waits for your yes), idempotent (re-running converges), and merge-safe (an existing mcp.json / settings.json keeps its other entries). unsetup removes exactly what setup wrote and leaves your .mokata/ config untouched.

See also