Skip to main content
On this page

Multi-agent setup

adk ships native wrappers for four agents. The same content (skills, agents, MCPs, shared guidelines) lives in one repo; install.sh translates it to each agent’s expected format and location.

Capability matrix

Feature Claude Code Cursor Codex CLI Junie
Skills (8 polymorphic) ✓ full — ~/.claude/skills/adk-* (symlinks) ✓ via rules — ~/.cursor/rules/adk-*.mdc ⚠ via prompts — ~/.codex/prompts/adk-*.md ⚠ via guidelines — ~/.junie/guidelines.md pointer
Subagents (9 personas as agents) ✓ full — ~/.claude/agents/adk-agent-*.md ⚠ inlined into rules ✗ no separate subagent context ✗ no separate subagent context
Slash commands ✓ ~/.claude/commands/adk-*.md (8 commands) ⚠ Cursor’s / palette per rule ✗ invoke via codex --prompt <name> ✗ invoke via Junie chat with explicit prompt
MCP merging ✓ ~/.claude/settings.json ✓ ~/.cursor/mcp.json ✓ ~/.codex/config.toml [[mcp_servers]] ⚠ partial; IDE-version dependent. Snippet written to agents-junie/junie-mcp.json.snippet for manual paste
Hooks (PreToolUse / PostToolUse / SessionStart) ✓ wired via install.sh ✗ not supported ✗ not supported ✗ not supported
Skill auto-routing by description ✓ Claude router reads description (1536-char cap) ✓ Cursor’s “agent-requested” rule mode ✗ no router; invoke explicitly ✗ no router; invoke explicitly
paths: glob frontmatter for auto-attach n/a ✓ Cursor uses globs: field n/a n/a
Global guidelines pointer (AGENTS.md) ✓ ~/.claude/CLAUDE.md append ✓ ~/.cursor/rules/_adk.mdc always-rule ✓ ~/.codex/instructions.md append ✓ ~/.junie/guidelines.md append

Claude Code

Full feature parity. The recommended primary agent.

  • Skills installed to ~/.claude/skills/adk-* (symlinks to <repo>/skills/adk-*).
  • 9 subagents (code-reviewer, security-reviewer, investigator, doc-writer, doc-reviewer, implementer, test-engineer, context-gatherer, explainer) installed to ~/.claude/agents/.
  • 8 slash commands at ~/.claude/commands/adk-{implement,review,investigate,document,sync,setup,improve,explain}.md.
  • All 9 MCPs merged into ~/.claude/settings.json (RAG skipped unless RAG_MCP_URL is set).
  • Hooks merged with _adk_managed: true marker so they’re idempotently updatable and removable.
  • ~/.claude/CLAUDE.md appended with a pointer to <repo>/AGENTS.md (idempotent, marker-based).

Cursor

Solid coverage. Native .mdc rule format maps well to skills.

  • 8 rules at ~/.cursor/rules/adk-*.mdc (symlinks); each rule has a description: and globs: for auto-attach.
  • One always-rule at ~/.cursor/rules/_adk.mdc that points at AGENTS.md.
  • All 9 MCPs merged into ~/.cursor/mcp.json.
  • Caveats: Cursor doesn’t have a separate subagent context — shared/personas/* are inlined into the rules where relevant. Cursor doesn’t have a hooks system equivalent to Claude Code’s, so the constitution is enforced by advisor prose only when running in Cursor.

Codex CLI

Partial. Codex’s plugin/skill ecosystem is less mature than Claude or Cursor; we ship what works today and surface gaps honestly.

  • 8 prompts at ~/.codex/prompts/adk-*.md — invoke via codex --prompt adk-implement (or whatever Codex’s prompt-invocation syntax becomes).
  • All 9 MCPs appended as [[mcp_servers]] blocks in ~/.codex/config.toml.
  • ~/.codex/instructions.md appended with the AGENTS.md pointer.
  • Gaps: no skill auto-routing by description; no separate subagent context; no slash commands; no --auto / --plan flag enforcement beyond what’s in the prompt body.

See agents-codex/README.md in the repo for the up-to-date capability table.

Junie (JetBrains AI)

Shallow. Junie’s MCP and rule story varies by IDE version.

  • ~/.junie/guidelines.md appended with the AGENTS.md pointer.
  • agents-junie/junie-mcp.json.snippet written; paste manually into Junie settings.
  • Gaps: no skill auto-routing; no separate subagent context; MCP wiring is manual; hooks not supported.

See agents-junie/README.md for the up-to-date list of what works.

When to use which agent

Task Best agent Why
Multi-file refactor with strict edit-format discipline Claude Code Hooks enforce constitution; plan/act tool restriction is real; subagents stay in clean contexts
Quick inline edits in your IDE Cursor globs:-based auto-attach surfaces the right rule for the file you’re in
CI / non-interactive context Codex CLI Headless invocation; no IDE required
JetBrains IDE work Junie Native to the IDE; limited adk integration but usable

You can run multiple agents in parallel. git pull in ~/code/agents-devkit updates all of them at once.

Adding a new agent

If a new agent (Aider, OpenHands, Continue, etc.) comes along, add agents-<env>/ and update install.py to handle it. The shared content in shared/ doesn’t need to change — only the per-env translation layer.

Next