The Agentuity Coder plugin adds a team of specialized AI agents, persistent memory, and cloud service access to Claude Code. It runs as a native Claude Code plugin with event hooks, slash commands, and auto-activated skills.
Installation
# Via Agentuity CLI (recommended)
agentuity ai claude-code install
# Via Claude Code marketplace
/plugin marketplace add agentuity/sdk
/plugin install agentuity-coder@agentuityRequires Claude Code, Agentuity CLI (agentuity auth login), and Bun. The plugin works without the CLI, but memory and cloud features will be unavailable.
Slash Commands
| Command | Description |
|---|---|
/agentuity-coder | Run a task with the full agent team |
/agentuity-cadence | Start autonomous long-running task execution |
/agentuity-cadence-cancel | Cancel an active Cadence session |
/agentuity-memory-share | Share content via Agentuity Cloud Streams |
/agentuity-sandbox | Run code in an isolated sandbox |
/agentuity-coder implement dark mode for the settings page
Agents also activate automatically based on context. You don't always need a slash command.
Agents
Seven agents with distinct roles, each running on a model tier suited to their task:
| Agent | Role | Model |
|---|---|---|
| Lead | Orchestrator: plans, delegates, synthesizes | opus |
| Scout | Explorer: codebase research, read-only | haiku |
| Builder | Implementer: code changes, tests, builds | sonnet |
| Architect | Autonomous implementer: complex multi-file work | opus |
| Reviewer | Code reviewer: catches issues, verifies quality | sonnet |
| Memory | Context manager: stores/recalls across sessions | haiku |
| Product | Requirements owner: PRDs, feature planning | sonnet |
Lead handles delegation automatically. For most tasks, describe what you want and the right agents are chosen for you.
Skills
The plugin includes auto-activated skills that inject Agentuity SDK expertise when relevant:
| Skill | Covers |
|---|---|
| agentuity-backend | @agentuity/runtime, @agentuity/schema, @agentuity/drizzle, @agentuity/postgres, @agentuity/evals |
| agentuity-frontend | @agentuity/react, @agentuity/auth, @agentuity/frontend, @agentuity/workbench |
| agentuity-ops | CLI commands, cloud services, deployments |
| agentuity-cloud | Package routing, ecosystem overview |
| agentuity-command-runner | Runtime detection, build/test/lint execution |
Memory
Persistent context across sessions via Agentuity Cloud:
- KV Storage for structured data (patterns, decisions, corrections)
- Vector Storage for semantic search over session history
- Entity-centric: tracks users, orgs, projects, repos
- Branch-aware: scopes memories to git branch context
Memory is saved automatically at session end and during Cadence iterations.
Cadence Mode
Cadence runs the agent team autonomously across multiple iterations until a task is complete:
/agentuity-cadence build the auth system with OAuth, session management, and tests
How it works:
- A Stop hook intercepts session end and re-injects the task prompt
- Memory checkpoints at each iteration for recovery
- The loop continues until the agent signals completion or max iterations is reached (default: 50)
Use --max-iterations N to cap iterations (e.g., /agentuity-cadence --max-iterations 10 build the auth system).
Cancel with /agentuity-cadence-cancel or Ctrl+C.
Hooks
The plugin registers event hooks that run automatically:
| Script | Event | What It Does |
|---|---|---|
session-start.sh | SessionStart | Detects Agentuity project, org, user, and git context |
session-end.sh | SessionEnd | Saves session memory (immediate KV + async processing) |
block-sensitive-commands.sh | PreToolUse | Blocks access to secrets, API keys, auth tokens |
pre-compact.sh | PreCompact | Saves memory before context window compaction |
cadence-stop.sh | Stop | Keeps the Cadence loop running until task is done |
stop-memory-save.sh | Stop | Prompts memory save before session ends |
Permissions
The install script configures Claude Code permissions in ~/.claude/settings.local.json:
- Allowed:
agentuity cloud *andagentuity auth whoami * - Blocked:
agentuity cloud secrets *,agentuity cloud secret *,agentuity cloud apikey *,agentuity auth token *
Deny rules take precedence. The PreToolUse hook adds a second layer blocking sensitive commands before they reach the permission system.
Cloud Services
Agents can use any agentuity cloud subcommand:
| Service | Description |
|---|---|
| KV | Key-value storage for structured data |
| Vector | Semantic search over stored content |
| Storage | File upload/download |
| Sandbox | Isolated code execution environments |
| Database | Postgres via agentuity cloud db |
| SSH | Connect to deployments |
Next Steps
- OpenCode Plugin: Alternative plugin for OpenCode with background agents and tmux
- AI Commands: Other AI-related CLI commands
- Creating Agents: Build your first agent