Four AI Coding Agents Compared: Open Source Wins After Privacy Scandals

Recent privacy scandals involving closed-source AI coding tools prompted a reevaluation of open-source alternatives. Grok Build was found transmitting entire Git repositories including history even with "improve model" disabled, while ZCode created encrypted snapshots containing .git history that only the vendor could decrypt. Both vendors later disabled uploads, but data deletion remains unverifiable. This led the author to prefer tools with open-source clients, even when models run in the cloud.

Mainstream Open-Source Harness Tools

Codex CLI

Codex CLI

is OpenAI's official terminal coding agent (Apache-2.0, GitHub: https://github.com/openai/codex). It works with a ChatGPT subscription or API key. The author used gpt-5.3-codex extensively, citing the best value: Chat access is nearly unlimited and doesn't consume subscription quota. However, from late GPT 5.6 through GPT 6, three major issues emerged: (1) model capability lagged behind Opus 5.5; (2) quota shrank drastically, especially with GPT-6-Astra where tokens drained rapidly; (3) speed dropped to ~20 tok/s, making it unusable for production. The author cancelled subscriptions, keeping only a basic plan to observe. OpenAI's quota resets don't resolve the fundamental compute-vs-demand imbalance. A practical note: new Codex CLI versions run a persistent background daemon; disable it via daemon_auto_start = false in [features] or launch with --no-daemon.

Codex's ecosystem is comprehensive—Chat, Work, Codex; CLI, Desktop, Dots—covering almost every AI tool form factor on the market.

Pi Agent

Pi

is a minimal open-source harness (MIT, https://pi.dev/). Its core is tiny: only read, write, edit files, and bash by default; everything else comes via extensions and skills. The author uses Pi most often but combines it with other tools for specific needs (e.g., Codex Desktop for previews, Codex CLI for Computer Use). Pi suits developers who: need to switch among multiple models quickly; want to change models mid-session to avoid quota exhaustion; value open-source, customizability, and token efficiency; want granular control over context and compression per model; and have custom workflows they don't want constrained by built-in agent tools. The author's custom Pi UI is available at https://github.com/flyeric0212/pi-packages.

Pi emphasizes minimalism, open source, and high customizability. Starts fast, opens instantly, switch models and go.

OpenCode

OpenCode

is an open-source AI coding agent (MIT, https://opencode.ai/) with terminal, IDE, and desktop clients. It integrates 75+ model providers and supports ChatGPT/GitHub Copilot account login. The author adopted it during DeepSeek-V4-Flash-0731's free period, which dramatically improved perceptions of domestic models. After two weeks DeepSeek raised prices. When GPT 5.6 introduced tiered pricing with covert hikes and quota cuts, the author diversified into Pi and OpenCode to reduce single-vendor dependency. OpenCode V2 added a default persistent background service; though disableable, the author rejected it and rolled back to V1. Codex CLI also added a daemon. The author argues CLI tools should exit cleanly; persistent daemons are unnecessary for most users and hard for non-experts to manage.

OpenCode works as a high-volume temporary supplement, chosen per task difficulty. Drawbacks: pricing follows external suppliers and changes without notice; quality fluctuates—backend may be official models, quantized versions, or self-hosted. Not recommended for long-term subscriptions; subscribe on demand based on reputation.

DeepSeek Harness

DeepSeek Harness

is DeepSeek's official open-source agent framework (MIT, GitHub: https://github.com/deepseek-ai/deepseek-harness). Everything is a plugin: models, tools, skills, UI. Early web-only; now has a desktop version. The author uses it less, mostly testing new releases. It's clean, fast to start, responsive, and feature-complete—a pure productivity tool. Hope it stays lean unlike Codex Desktop's bloat. If DeepSeek launches a Coding Plan, adoption would rise. Currently API pay-as-you-go with peak/off-peak pricing: weekdays 9:00-12:00 and 14:00-18:00 are peak (2x price), exactly when domestic developers code most. Off-peak only helps nights/weekends—unfriendly to regular developers. Yet model speed is ~300 tok/s; after experiencing that, 20 tok/s feels intolerable. With clear requirements and good references, results can be surprisingly good.

DeepSeek Harness shows strong potential; a Coding Plan would attract more users. Peak pricing aligns with domestic work hours, making it costly for typical developers. But 300 tok/s speed is a game-changer—once accustomed, 20 tok/s is painful.

Four-Tool Comparison

Codex CLI

Positioning: OpenAI official terminal coding agent

License: Apache-2.0

Model swappable: Default OpenAI; custom providers via Responses protocol

Learning curve: Low, subscribe and log in

Best for: Existing ChatGPT subscribers wanting full ecosystem

Pi Agent

Positioning: Minimal agent harness, build via extensions

License: MIT

Model swappable: Yes, 15+ providers, API Key/OAuth, switch mid-session

Learning curve: High, configure extensions and permissions yourself

Best for: Custom workflows, context/token control

OpenCode

Positioning: Open-source coding agent, terminal/IDE/desktop

License: MIT

Model swappable: Yes, 75+ providers, ChatGPT/Copilot accounts

Learning curve: Low, install and use

Best for: Avoid vendor lock-in, out-of-the-box high volume

DeepSeek Harness

Positioning: General agent harness, everything pluggable, developer preview

License: MIT

Model swappable: Yes, DeepSeek plus third-party API keys and custom gateways; OAuth not yet

Learning curve: Medium, requires API key setup

Best for: Building your own agent, accepting preview status

Closing Thoughts

Prefer open-source tools—they're at least transparent.

Define your workflow first, then pick tools that fit. Others' recommendations may not suit you.

Switch AI subscriptions freely. Model capabilities evolve fast; options multiply. Mix and match instead of locking into one vendor.