Single source of truth for 6 AI coding agents:
- OpenCode
- Codex (ChatGPT)
- Claude Code
- GitHub Copilot
- Hermes
- Gemini CLI
Just talk to it. Give agent way to verify. Ship fast.
# One-liner (no deps)
curl -fsSL https://raw.githubusercontent.com/hugocarreira/agentrc/main/cli/install.sh | bash
agentrc setup
# Or tell your AI: "install the harness from github.com/hugocarreira/agentrc"agentrc setup # Link global config + skills (one-time)
agentrc init <n> # Create AGENTS.md + FEATURES.yaml + PROGRESS.md
agentrc link <n> # Reference existing project
agentrc verify # Check global setup + current project pack
agentrc status # Overview of all agents and config
agentrc guide # Friendly walkthroughagentrc/
├── AGENTS.md RTK.md SETUP.md
├── cli/ # Go CLI + install.sh
├── skills/ # 16 focused workflows
└── plugins/ # OpenCode plugins
Agent reads: harness config → global rules → project rules → current state → your code.
agentrc init creates:
AGENTS.mdfor project-specific rules, stack, commands, and boundariesFEATURES.yamlfor active feature scope and acceptance criteriaPROGRESS.mdfor handoff notes, blockers, and exact restart point
FEATURES.yaml keeps feature state lightweight:
features:
- id: pricing-page
behavior: Site has a pricing page with monthly and yearly plans
done_when:
- /pricing route renders
- toggle changes displayed prices
- primary CTA links to signup
check:
type: manual
note: Run app and verify in browser
state: activeUpdate PROGRESS.md at the end of substantial sessions. Update FEATURES.yaml only when feature state, scope, or acceptance changes.
- Give Agent Way to Verify — Tests, screenshots, benchmarks
- Context is Precious —
/clearoften,/btwfor quick questions - Just Talk To It — Short prompts + screenshots, no theater
- Final Manual Review — Always understand what agent did
16 skills · 6 agents · 1 source of truth · 0 theater