Small crew. Sharp work.
OpenCode-first agent orchestration for planning, building, UI polish, review, and security-aware checks.
Install · Uninstall · Commands · Crew · Architecture · Models
AutoCrew is a small, reusable crew of coding-agent prompts. It gives your assistant enough structure to plan, delegate, build, review, and ship without dragging in a heavyweight multi-agent runtime.
The main agent stays accountable. Specialist agents are called only when they add real value: repo scouting, implementation, review/security, or UI/UX design quality. The result is stronger work with lower context noise.
| Built for | Avoids |
|---|---|
| OpenCode-native agents and slash commands | Heavy npm package or daemon |
| Global or project-local setup | Provider wizard and hard-coded models |
| Claude Code-compatible files | Tool lock-in |
| UI/UX design specialist | Generic AI-looking frontend work |
| Reviewer with security checks built in | Separate security-command sprawl |
| Compact file-first helper output | Long subagent transcripts |
| Caveman/Karpathy-style prompt discipline | Required external skills |
AutoCrew is inspired by orchestration projects such as oh-my-openagent, but it is not a fork and does not copy upstream code, agents, team runtime, telemetry, provider setup, or packaging.
It also folds in prompt-level infrastructure ideas from Caveman and Andrej Karpathy Skills: compact communication, file-first outputs, think-before-coding, simplicity-first implementation, surgical diffs, and verify-before-claiming-done behavior.
Recommended: let your coding agent install it. Copy this prompt into OpenCode, Claude Code, or another capable coding agent:
Install AutoCrew from https://github.com/leobbaroni/AutoCrew.
Read `AI_INSTALL.md` from that repo before doing anything.
Follow it exactly.
Before copying files, ask me:
1. OpenCode, Claude Code, or both?
2. Global install or project-local install?
3. Use default models or configure specific models per agent?
Recommend the best option for my setup, then install and verify it.
AI_INSTALL.md gives the installing agent the exact file map, questions to ask, recommended defaults, copy rules, model policy, and verification checklist.
| Target | Scope | Destination |
|---|---|---|
| OpenCode | Global | ~/.config/opencode/agents/, ~/.config/opencode/commands/ |
| OpenCode | Project | .opencode/agents/, .opencode/commands/ |
| Claude Code | Global | ~/.claude/agents/, ~/.claude/commands/ |
| Claude Code | Project | .claude/agents/, .claude/commands/ |
Recommended default: OpenCode + global install + default models.
PowerShell:
git clone https://github.com/leobbaroni/AutoCrew.git
cd AutoCrew
$paths = opencode debug paths
$config = ($paths | Select-String '^config\s+(.+)$').Matches[0].Groups[1].Value.Trim()
New-Item -ItemType Directory -Force "$config\agents", "$config\commands" | Out-Null
Copy-Item .opencode\agents\*.md "$config\agents\" -Force
Copy-Item .opencode\commands\*.md "$config\commands\" -Force
opencode agent listmacOS, Linux, and WSL:
git clone https://github.com/leobbaroni/AutoCrew.git
cd AutoCrew
mkdir -p ~/.config/opencode/agents ~/.config/opencode/commands
cp .opencode/agents/*.md ~/.config/opencode/agents/
cp .opencode/commands/*.md ~/.config/opencode/commands/
opencode agent listProject-local OpenCode:
cp -R path/to/AutoCrew/.opencode .AutoCrew has no package manager state, runtime, database, or background service. Uninstalling means removing the copied agent and command files from the install target you chose.
Recommended: let your coding agent remove it safely. Copy this prompt:
Uninstall AutoCrew.
Read `AI_INSTALL.md` from https://github.com/leobbaroni/AutoCrew before doing anything.
Follow the uninstall section exactly.
Before deleting files, ask me:
1. Was AutoCrew installed for OpenCode, Claude Code, or both?
2. Was it installed globally or project-locally?
3. Should any customized AutoCrew agent files be backed up first?
Remove only AutoCrew files, verify they are gone, and summarize exact paths touched.
Manual uninstall targets:
| Target | Scope | Remove |
|---|---|---|
| OpenCode | Global | ~/.config/opencode/agents/crew-*.md, ~/.config/opencode/commands/crew*.md |
| OpenCode | Project | .opencode/agents/crew-*.md, .opencode/commands/crew*.md |
| Claude Code | Global | ~/.claude/agents/crew-*.md, ~/.claude/commands/crew*.md |
| Claude Code | Project | .claude/agents/crew-*.md, .claude/commands/crew*.md |
If you customized any AutoCrew files after installing, back them up before removal.
/crewplan redesign the billing settings flow
/crewwork add optimistic updates to the kanban board
| Command | Behavior |
|---|---|
/crewplan |
Plans only. Inspects enough context to produce a decision-complete implementation plan. |
/crewwork |
Executes. Inspects, delegates only when useful, edits, verifies, and summarizes. |
| Agent | Job | Called when |
|---|---|---|
crew-lead |
Orchestrates scope, delegation, integration, verification, and final response. | Always through /crewwork or /crewplan. |
crew-scout |
Finds files, source-of-truth docs, entry points, patterns, tests, and conventions. | The repo is unfamiliar or the relevant surface is unclear. |
crew-worker |
Makes one bounded implementation change and verifies it. | The task has a clear write scope. |
crew-reviewer |
Finds bugs, regressions, weak tests, and security risks. | The change is risky, security-sensitive, or ready for review. |
crew-designer |
Raises UI/UX quality, accessibility, product copy, and visual polish. | The task touches frontend, layout, product flows, or interaction design. |
The crew is intentionally small. Performance matters more than token shaving, but every helper has a compact output contract so the lead gets signal instead of transcripts.
AutoCrew is just Markdown configuration:
.opencode/agents/*.md
.opencode/commands/*.md
.claude/agents/*.md
.claude/commands/*.md
AI_INSTALL.md
There is no runtime, registry, telemetry, background process, or dependency install. Copy the files, use the commands.
AutoCrew ships without model fields. Your tool defaults apply unless you choose otherwise.
If you want per-agent routing, add model fields to your installed copies:
model: anthropic/claude-sonnet-4-5Suggested routing:
| Agent | Model recommendation |
|---|---|
crew-lead |
Strong general coding/reasoning model |
crew-worker |
Strong coding model |
crew-reviewer |
Strongest review/reasoning model |
crew-designer |
Strong UI/UX and visual-reasoning model |
crew-scout |
Fast model is fine for simple lookup; use stronger model for architecture discovery |
AutoCrew bakes these rules into the agents:
- Inspect source truth before guessing.
- Keep implementation simple and scoped.
- Touch only the assigned area.
- Use locate, fix, verify loops.
- Treat design quality as product quality.
- Review by severity, not preference.
- Keep security findings practical and bounded.
- Return compact, file-first helper output.
- No npm package.
- No telemetry.
- No provider setup.
- No large named-agent roster.
- No copied upstream code.
- No mandatory external skills.
- No standalone security command.
- No hard-coded models.
MIT. See LICENSE.