This is the beginner’s gateway to AI coding agents. Every setup, installation, and getting-started article links here. If you’ve never used a coding agent, start here.
Sister pillars: Best coding agents guide · Security checklist · Pricing guide
What is a coding agent?
A coding agent is an AI tool that writes, edits, reviews, and sometimes runs code for you — directly in your terminal or IDE. Unlike a chatbot, it has access to your filesystem, git, and sometimes your browser. It can refactor entire codebases, run tests, create PRs, and automate repetitive tasks.
Key difference from autocomplete: An autocomplete suggests the next line. A coding agent can plan a multi-file refactor, execute it, run tests, and fix what breaks.
Quick start: pick your path
| Your situation | Start here | Time to first success |
|---|---|---|
| I want the easiest setup | Cursor | 5 minutes |
| I want free + powerful terminal | Hermes or OpenCode | 10 minutes |
| I want the deepest reasoning | Claude Code | 10 minutes |
| I want GitHub integration | Copilot CLI | 5 minutes |
| I want cloud parallelism | Codex | 10 minutes |
| I want local-first, no API costs | Run with Ollama | 20 minutes |
Agent-by-agent setup guides
Claude Code
Claude Code is Anthropic’s terminal-based coding agent. It offers the deepest reasoning but requires a subscription or API key.
Prerequisites:
- Node.js 18+ installed
- Terminal (Mac, Linux, or Windows with WSL/Git Bash)
- Anthropic API key or Claude subscription
Installation:
npm install -g @anthropic-ai/claude-codeFirst run:
claude
# Follow the auth flow — paste your API key
# Navigate to your project directory
cd your-project
claudeRecommended next steps:
- Best coding agent setup after 6 months — proven configurations
- AGENTS.md: the one trick that makes agents 10x smarter — create an AGENTS.md in your repo
- Claude Code: Skills vs Subagents vs MCP — understand the extension model
- Claude Code Deep Dive — architecture and advanced usage
Windows users: See Claude Code on Windows — click-through permissions and Windows hooks and shell redirection for known issues and fixes.
Cursor
Cursor is an IDE-based coding agent (fork of VS Code). It’s the easiest entry point if you’re coming from a traditional editor.
Prerequisites:
- Download from cursor.com
- Git installed
Setup:
- Download and install Cursor
- Open your project folder
- Enable the AI features in Settings → AI
- Use Cmd/Ctrl+K for inline edits, Cmd/Ctrl+L for chat
Recommended next steps:
- Why Cursor Wins on Speed — understand the advantage
- Cursor’s Background Agents — set up autonomous background tasks
- Is Cursor Worth It in 2026? — community verdict
Hermes Agent
Hermes is the most feature-rich open-source coding agent — cron jobs, multi-provider routing across 15+ LLMs, browser automation, and memory. Free and self-hosted.
Prerequisites:
- Python 3.10+ or Node.js
- Terminal (works on Windows, Mac, Linux)
- API key for at least one LLM provider (or use OpenRouter for access to many)
Installation:
# Clone and install
git clone https://github.com/NousResearch/hermes-agent.git
cd hermes-agent
pip install -e .
# Or use the installer
curl -fsSL https://hermes.nousresearch.com/install.sh | bashFirst run:
hermes
# Follow the setup wizard
# Configure your API provider
# Start codingRecommended next steps:
- Hermes Agent Deep Dive — full architecture overview
- Hermes v0.19: 80% faster startup — what changed recently
- Extending Hermes with Custom Skills — plugin system
- Setting Up Autonomous Cron Jobs — automation
- Hermes Security: Credential Guards — secure your setup
- Local-First Architecture — privacy advantages
OpenCode
OpenCode is a skill-driven, provider-neutral open-source coding agent. BYO keys, no vendor lock-in.
Prerequisites:
- Go 1.21+ installed
- Terminal
- API key for your preferred provider
Installation:
# Install via Go
go install github.com/opencode-ai/opencode@latest
# Or download from GitHub releasesRecommended next steps:
- OpenCode Deep Dive — architecture
- Why Provider-Neutral Design Is Its Biggest Advantage — flexibility
- OpenCode + GitHub Copilot provider — use Copilot keys
OpenAI Codex
Codex is OpenAI’s cloud-parallel coding agent. It runs in cloud sandboxes and can execute multiple tasks simultaneously.
Prerequisites:
- ChatGPT Plus subscription ($20/month)
- Terminal
Setup:
npm install -g @openai/codex
codex --auth
# Follow the auth flowRecommended next steps:
- Codex Deep Dive: Parallel Agent Future — architecture
- Codex Cloud Execution Economics — cost analysis
- Codex Sandbox Is Dead on Windows — Windows users beware
GitHub Copilot CLI
Copilot CLI is GitHub’s terminal agent, deeply integrated with GitHub repos, PRs, and Actions.
Prerequisites:
- GitHub account with Copilot access ($10-39/month)
- Terminal
Setup:
# Install
npm install -g @githubnext/copilot-cli
# Authenticate
copilot authRecommended next steps:
- Copilot CLI: Terminal Agent That Knows Your Repos
- Why I Stopped Using Copilot and Went Full Terminal Agent
Other agents worth trying
- Kilo Code CLI: Lightweight Agent for Quick Tasks — minimal, fast
- Codebuff: Terminal Agent That Does One Thing Well — focused simplicity
- Oh My Pi Deep Dive — model exploration and experimentation
- pi.dev: Agent That Remembers Everything — knowledge-backed long runs
- Gitlawb Zero Deep Dive — decentralized, you own everything
- Goose: Universal Coding Agent via MCP — tool integration
- Grok Build: xAI’s Terminal Coding Agent — newcomer from xAI
Running with Ollama (local models)
If you want zero API costs, you can run coding agents with local models via Ollama.
Prerequisites:
- Ollama installed (ollama.com)
- 8GB+ RAM (16GB+ recommended for good models)
- Compatible coding agent (OpenCode, Hermes, Oh My Pi, Kilo)
Full guide: How to Run Coding Agents with Ollama: The Complete Local Setup Guide
Key considerations:
- Quality depends heavily on model size — 7B models are functional, 70B models are competitive
- Speed depends on your hardware — GPU acceleration helps significantly
- Context windows are smaller than cloud APIs
- Good for learning, experimentation, and privacy-sensitive code
Running in the browser
Some developers prefer not to install anything locally. The browser-remote trend is growing.
Full guide: Run Claude Code and Codex in Your Browser — The Browser-Remote Trend Explained
Making any agent 10× better
The harness matters less than the instructions you feed it. Two files can transform any agent:
AGENTS.md — the universal instruction sheet
An AGENTS.md file in your repo root tells the agent how your project works — coding style, test commands, architecture decisions, things to avoid.
- The AGENTS.md File: The One Trick That Makes Every Coding Agent 10x Smarter
- AGENTS.md Complete Guide
UI.md — the design rulebook
Context engineering
- Context Engineering for Coding Agents: How to Make Every Token Count
- Context Compression Middleware for AI Coding
- The Brevity Revolution: How Shutting Up Makes Agents Smarter
Security before production
Before you wire an agent into a real project with real secrets:
- Run the security checklist — Coding Agent Security Checklist 2026
- Check your sandbox — make sure it’s actually isolated
- Audit your .env files — the agent can probably read them
- Set up permission gates — never auto-approve destructive operations
Common setup mistakes to avoid
Before you start, here are the pitfalls that trip up most beginners:
-
Skipping AGENTS.md. The single biggest upgrade you can make. Create an
AGENTS.mdfile in your repo root before you run any agent for the first time. Why it matters. -
Not setting token limits. Without limits, a single bad loop can burn $50+ in API costs. Always set
--max-tokensor equivalent. See the pricing problem nobody talks about. -
Running in production repos without sandbox verification. Test any agent in a throwaway project first. Verify the sandbox actually works. Many don’t.
-
Ignoring your operating system. Windows users face unique issues. Codex sandbox is dead on Windows. Claude Code has permission bugs. Check the Windows-specific articles before blaming the agent.
-
Not committing before the agent runs. Always
git stashor commit before giving an agent access. Background tasks can corrupt git state. -
Expecting too much too soon. Agents are fastest when they understand your codebase. The first session is always slower than the tenth. Give it time to build context.
What to read next
- The Complete Guide to AI Coding Agents in 2026 — full landscape
- Coding Agent Features Comparison 2026 — feature matrix
- Best Coding Agents 2026 — Complete Guide — decision framework
- What Coding Agents Actually Cost Per Month — budget planning
- Best Coding Agent Setup After 6 Months — proven configurations
- Context Engineering for Coding Agents — make every token count
Cross-pillar navigation
Last updated: July 27, 2026. When a new agent launches or setup instructions change, we update this guide.