Overlord's Guides
REFERENCE // HARNESS MATRIX

Code Harnesses

The agentic AI coding tools covered across Overlord’s Guides. Every directive is harness-agnostic — the patterns transfer. This is the deep reference; the Start Here page has the short version and a decision tree.

The Overlord’s disclosed preference

The Overlord’s favorite is OpenCode — open-source, local-first, agent-agnostic. Disclosed openly, because the Overlord has no financial interest and a strong correct one. The patterns you learn here transfer to whichever harness you switch to later. That transferability is the entire premise of these guides.

OpenCode

Overlord's pick

Open-source, local-first, agent-agnostic. You bring the model; OpenCode brings the harness.

opencode.ai ↗

OpenCode is an open-source (MIT) agent harness designed around a single principle: you own the stack. It does not bundle a model, it does not phone home, and it does not make decisions about your tools for you. You wire the MCP servers, you pick the model, you set the permissions. In return you get a harness that is transparent, auditable, and portable across every model provider you ever switch to.

Strengths

  • + Open-source (MIT) — fork it, audit it, contribute back
  • + Local-first: no telemetry lock, no vendor data pipeline
  • + Agent-agnostic: bring Claude, GPT, local models, anything with an API
  • + Full control over MCP servers, tools, permissions, and hooks
  • + Free to run; you only pay the model provider you already chose
  • + Ships with some free model inference out of the box; subscribe to OpenCode Go for cheap hosted models
  • + Patterns learned here transfer to any other harness

Tradeoffs

  • − More setup — you wire MCP servers and model routes yourself
  • − Docs assume you can read a config file (you can)
  • − No bundled model — bring your own API key, local inference, or an OpenCode Go subscription
  • − Less hand-holding than the proprietary options
License: MIT (open-source) Model: Bring your own (any provider) + free inference + OpenCode Go Cost: Free software + your model costs (free inference included; OpenCode Go for cheap hosted models)
ACP: Yes (dedicated ACP support) Rules file: AGENTS.md (cross-harness standard) Skills: .agents/skills/ with SKILL.md (agentskills.io)
Best for: Devs who want to own the stack end to end. Pick if: You want to own your agent stack, you are comfortable with config, and you would rather understand the tool than be understood by it.

Claude Code

Popular

Anthropic's terminal-native agent. Polished, opinionated, strong reasoning out of the box.

www.anthropic.com/claude-code ↗

Claude Code is Anthropic's terminal-native coding agent. It is the most polished guided experience in the field: install it, point it at a repo, and it runs. The reasoning quality on Claude models is strong, the tool-use behavior is mature, and the opinionated defaults are mostly right for common workflows. The tradeoff is the opinion: you adapt to Claude Code's workflow, not the other way around, and you are locked to Anthropic models and Anthropic pricing.

Strengths

  • + Strong reasoning on Claude models, especially long-context and tool-use
  • + Terminal-native: low friction, install and go
  • + Opinionated defaults that are mostly right for common workflows
  • + Plan mode and tool-use behavior are the most mature in the field

Tradeoffs

  • − Proprietary — Anthropic-locked, not self-hostable
  • − Cost scales with usage; no bring-your-own-model escape valve
  • − "Opinionated" means opinionated — you adapt to it, not vice versa
  • − Local models via Ollama are technically possible but unsupported as a standard configuration
License: Proprietary Model: Claude (Anthropic only); local models via Ollama possible but non-standard Cost: Anthropic API usage (pay per token)
ACP: Yes (via Zed SDK adapter) Rules file: CLAUDE.md + .claude/rules/ Skills: .claude/skills/ with SKILL.md (agentskills.io)
Best for: Devs who want a polished, guided experience today. Pick if: You want it to just work, today, in your terminal; you are already in the Anthropic ecosystem; and "it just works" is worth more to you than "I can change anything."

Codex

Popular

OpenAI's code-generation agent. Fast, broad language coverage, fits the OpenAI stack.

openai.com/codex ↗

Codex is OpenAI's code-generation agent. It is fast, covers a broad range of languages and frameworks, and integrates cleanly with the rest of the OpenAI ecosystem — API, playground, evals, fine-tuning. It is strongest as a generate-code-on-demand tool rather than a run-an-autonomous-loop tool: you invoke it, it produces code, you integrate it. If your stack is already OpenAI-shaped, that is a feature, not a bug.

Strengths

  • + Fast iteration on OpenAI models (GPT and o-series)
  • + Broad language and framework coverage out of the box
  • + Integrates with the OpenAI ecosystem: API, playground, evals, fine-tuning

Tradeoffs

  • − Proprietary — OpenAI-locked
  • − OpenAI pricing and rate limits apply
  • − Less "run an autonomous loop" than the dedicated agent harnesses — more "generate code, you integrate it"
  • − Local models via Ollama are technically possible but not a standard configuration and may require additional setup
License: Proprietary Model: OpenAI (GPT / o-series); local models via Ollama possible but non-standard Cost: OpenAI API usage (pay per token)
ACP: Yes (via Zed SDK adapter) Rules file: AGENTS.md (cross-harness standard) Skills: Skills & Plugins system (build-skills)
Best for: Devs who already live in OpenAI tooling. Pick if: You already use OpenAI tooling, you want code generation that fits that stack, and you do not need a fully autonomous loop.

Antigravity

Popular

Google's agentic coding tool. Gemini-powered, integrated into the Google ecosystem.

antigravity.google/ ↗

Antigravity is Google's agentic coding tool, powered by Gemini models and integrated into the Google developer ecosystem. It supports a range of coding workflows from inline assistance to autonomous task execution. As a Google product, it benefits from Google's infrastructure, model investment, and ecosystem integration. The product is actively developed and evolving; consult the official documentation for current capabilities and pricing before committing.

Strengths

  • + Backed by Google — infrastructure, model lineup, and long-term runway
  • + Powered by Gemini models with deep Google ecosystem integration
  • + Supports a range of workflows from inline assistance to autonomous execution

Tradeoffs

  • − Proprietary — Google-locked, tied to the Google stack
  • − Still evolving rapidly; expect changes to surfaces and capabilities
  • − Ecosystem lock-in: the more you integrate, the more Google-shaped your stack becomes
License: Proprietary Model: Google (Gemini) Cost: Varies (check the official docs for current terms)
ACP: No Rules file: GEMINI.md (global) + .agents/rules/ (workspace) Skills: .agents/skills/ with SKILL.md (agentskills.io)
Best for: Devs who already live in the Google ecosystem or want a major-vendor option. Pick if: You are already in the Google stack, you want a major-vendor-backed harness, and you are comfortable reading the official docs to keep up with a fast-moving product.

Cursor

Popular

The AI-native editor. Familiar shell, AI baked into the muscle memory.

cursor.com ↗

Cursor is an AI-native code editor built on a VS Code fork. The pitch is simple: it is the editor you already know, with an agent layer woven into the surface instead of bolted on. You get inline completions, chat, and an agent mode that can edit files and run commands in the open workspace. For devs who never wanted to leave the editor to talk to an agent, this is the lowest-friction on-ramp in the field. The tradeoff is the editor: you are in Cursor’s world, on Cursor’s pricing, and the agent loop is bounded by what the editor surface exposes.

Strengths

  • + Lowest friction of any harness — it is just your editor, with an agent
  • + Inline completions + chat + agent mode in one surface
  • + Familiar to anyone who already lives in VS Code
  • + Fast iteration loop; great for in-flow refactors and edits

Tradeoffs

  • − Proprietary editor — you are on Cursor’s pricing and roadmap
  • − Agent loop is bounded by the editor surface; less of a bare-terminal loop than OpenCode or Claude Code
  • − Less control over MCP servers, hooks, and permissions than the harness-first tools
License: Proprietary Model: Bring your own (Anthropic, OpenAI) + Cursor-hosted options Cost: Free tier + Pro/Business subscription
ACP: Yes (native ACP via Cursor CLI) Rules file: .cursor/rules/ (MDC files) Skills: Docs at docs.cursor.com/context/skills
Best for: Devs who want the agent inside the editor they already use. Pick if: You live in VS Code, you want the agent to meet you where you already are, and "it just works inside the editor" is worth more to you than "I own the entire agent stack."

GitHub Copilot

Popular

GitHub's AI coding agent. IDE-native, cloud agent, CLI — sits where you already work.

github.com/copilot ↗

GitHub Copilot is GitHub's AI coding product. It is not a single surface — it is a suite that spans inline IDE completions, Copilot Chat, a cloud agent that opens PRs from a task description, a CLI, and a desktop app. If your team already lives in GitHub (and most do), the on-ramp is nearly frictionless: the agent, the code review, the PR lifecycle, and the repository context are all in the same place you already authenticate. The tradeoff is the ecosystem: you are in GitHub’s world, on GitHub’s plans and policies, and the agent is bounded by what the GitHub surfaces expose.

Strengths

  • + IDE-native across VS Code, JetBrains, Visual Studio, Xcode, and Eclipse
  • + Cloud agent runs tasks and opens PRs directly from GitHub
  • + CLI for terminal-driven agent sessions with fleet, research, and automations
  • + Free tier available; Pro/Pro+/Max upgrades unlock more models and limits
  • + Deep GitHub integration — issues, PRs, code review, and agent in one place
  • + Bring-your-own-key (BYOK) supported for external model providers

Tradeoffs

  • − Proprietary — GitHub-locked, bounded by GitHub’s surfaces and policies
  • − Plan and usage limits apply; heavy agent use can hit AI-credit caps
  • − Agent loop is shaped by the GitHub workflow; less of a bare-terminal loop than OpenCode
License: Proprietary Model: GitHub-hosted (multi-model) + BYOK Cost: Free tier + Pro/Pro+/Max subscription (AI credits)
ACP: Yes (public preview) Rules file: .github/copilot-instructions.md + AGENTS.md Skills: Not confirmed (no skills directory documented)
Best for: Devs and teams who already live in GitHub and want the agent where they already work. Pick if: Your team already lives in GitHub, you want inline completions plus a cloud agent that opens PRs, and "the agent is where my code already lives" is worth more to you than owning the entire agent stack.

How the harnesses relate to the guides

The directives do not teach you OpenCode, or Claude Code, or Codex. They teach you how to operate an agent — any agent — without shipping yourself. The harness is the execution surface; the patterns are the skill. Pick whichever harness fits your constraints in the decision tree, then work through the directives. The Disaster Logs you file will look the same no matter which one you chose.

Start Here Browse directives