In 2026, CLI coding agents have reached maturity. The big three come from AI giants: Claude Code (Anthropic), Gemini CLI (Google), Codex (OpenAI). But the open source community is pushing hard with Hermes Agent (NousResearch) and OpenClaw (ex-Clawdbot). They all do the same thing, read your code, modify it, run commands in your terminal, but they do it differently.
After months of intensive use, here’s our honest comparison. No absolute ranking: the best tool depends on your context, your budget, and what you’re building.
All support MCP (Model Context Protocol), meaning they can connect to the same external tools. The real differences lie in reasoning quality, customization ecosystem, and pricing model.
Quick verdict
Claude Code is the strongest default for production engineering teams because its reasoning quality and operating harness (CLAUDE.md, Skills and Hooks) are the most mature. Gemini CLI is the best free entry point, Codex is the most permissively licensed option, and Hermes Agent is the most interesting choice for multi-model experimentation and persistent personalization. OpenClaw is powerful, but should only be deployed after a deliberate security review.
| If your priority is… | Start with | Main trade-off |
|---|---|---|
| Production code quality and team governance | Claude Code | Paid subscription and closed source |
| A capable agent with no subscription cost | Gemini CLI | Less mature customization ecosystem |
| Open-source integration and native subagents | Codex | Variable cost and less predictable output quality |
| Multi-model use and automatic personalization | Hermes Agent | Younger ecosystem |
| Persistent personal automation across apps | OpenClaw | Highest security and configuration burden |
How this comparison was evaluated
The verdict is based on repeated use in real repositories rather than a one-shot benchmark. We compared five dimensions that matter in production: quality on multi-file changes, context retention, tool and MCP integration, team governance, and total operating cost. Product capabilities and prices change quickly, so each dated update is checked against the official documentation and changelogs linked at the end of the article.
July 2026 update: what changed since April
The landscape moved fast in three months. The essentials, verified against each project’s official changelogs:
Hermes Agent changed category. The project that was three weeks old in April now shows around 15,000 GitHub stars and 370+ contributors. Version 0.18, the “Judgment Release” (July 1), brings two notable advances: Mixture-of-Agents ensembles became selectable models like any other (a “council” of frontier models deliberating, with each model’s reasoning visible), and the agent now verifies its work against evidence instead of declaring a task done. The team also cleared 100% of its critical backlog in twelve days. The learning loop that impressed us in April is now backed by a multi-surface platform (desktop, mobile, messaging apps) and in-house models (Hermes 4.3, focused on local intelligence).
Claude Code made Sonnet 5 its default model with a 1 million token context window (promotional pricing through late August 2026), erasing Gemini CLI’s main advantage. Subagents now run in the background by default and can open pull requests when they finish, and browser automation (Claude in Chrome) reached general availability. The default permission mode also got stricter, a clear governance-first signal.
Codex and Gemini CLI keep shipping fast: Codex is stacking Rust releases (0.144 in mid-July) and Gemini CLI iterates with near-daily nightlies. OpenClaw continues its rapid beta cycles. No major repositioning for these three: the strengths and weaknesses described below still hold.
What this changes in the verdict: the “1M tokens” argument is no longer exclusive to Gemini CLI, and Hermes has moved from experimental bet to credible ecosystem for personal, multi-surface use. For professional team usage, the ranking below stands.
The table
| Claude Code | Gemini CLI | Codex | Hermes Agent | OpenClaw | |
|---|---|---|---|---|---|
| Maker | Anthropic | OpenAI | NousResearch | Community (ex-steipete) | |
| Model | Opus / Sonnet | Gemini 2.5 Pro | GPT-5-Codex | Multi (OpenRouter, Gemini, local) | Multi (Kimi 2.5, GPT, Claude) |
| Context window | 1M tokens (Sonnet 5) | 1M tokens | Variable | Depends on model | Depends on model |
| Cost | Pro/Max ($20-100/mo) | Free (Google account) | Pay-per-use API | API fees only | Free + API fees |
| Open source | No | Yes (Apache 2.0) | Yes (MIT) | Yes (Apache 2.0) | Yes |
| MCP | Yes | Yes | Yes | Yes (OAuth 2.1) | Partial |
| Multi-agent | Teams | No | Native subagents | No | No |
| Hooks/customs | Hooks + Skills | Basic config | Plugins | Auto-generated skills | Community skills |
| Platforms | CLI + Desktop + Web | CLI | CLI | CLI + Telegram + Discord + Slack | CLI + WhatsApp + Slack |
| Memory | CLAUDE.md (manual) | No | No | Learning loop (auto) | Persistent local |
| Maturity | ~1 year (mature) | Recent | Recent | ~4 months, very active (15K stars) | Young but viral (60K stars) |
Already deploying agents in production ? A CLI is not a plugin. See our two turnkey offerings :
- Business AI Agent Plugins — skills + hooks + MCP packaged, tested, benchmarked, source code delivered. Built with the Anthropic-certified agentic architect methodology (CCA Foundations).
- AI Security Audits — pentest, threat modeling and vulnerability detection accelerated by AI agents. OWASP, MITRE ATT&CK, CIS frameworks. Covered by Colombani.ai professional liability insurance.
Claude Code : the most mature
Claude Code is the oldest and most polished of the three. Its main advantage: reasoning quality. Opus excels at complex tasks : multi-file refactoring, subtle bug hunting, system architecture.
The agentic loop. Claude Code sends a request to the Claude API. The response includes a stop_reason: "tool_use" means the model wants to use a tool (read a file, run a command), "end_turn" means it’s done. The model decides when to stop, not external logic.
Agentic search. Anthropic abandoned the RAG approach (vector database) in favor of iterative search: Grep, Glob, Read. The model searches, refines, searches again. Slower than a vector index, but significantly more reliable on real code.
Auto mode uses classifiers that evaluate the risk of each action before executing it. Writing a test file: low risk, automatic execution. Deleting a directory: high risk, asks for confirmation.
The harness. This is where Claude Code truly stands apart. The CLAUDE.md system provides persistent context to the model : coding conventions, project architecture, business constraints. Skills encapsulate reusable workflows. Hooks (PreToolUse, PostToolUse) let you inject logic before or after each action. Teams enables multi-agent work, worktrees enable parallel work on isolated branches.
The downsides. Claude Code is not open source. It requires a Pro or Max subscription ($20-100/month). Since July 2026, the historical context limitation is gone: Sonnet 5 ships as the default with a 1 million token window.
Gemini CLI : the most accessible
Gemini CLI is Google’s agent, released more recently. Its killer argument: it’s free with a Google account, and the context window is 1 million tokens.
1M tokens of context. Long a Gemini exclusive, this context level is now also Claude Code’s with Sonnet 5. In practice, you can load an entire codebase into context without the model losing track. For code exploration or understanding an unfamiliar project, this is a real advantage.
Google Search grounding. Gemini CLI can search the web natively. The model checks information in real time : recent documentation, latest library versions, Stack Overflow content. Other agents can do this via MCP, but here it’s built in.
Open source (Apache 2.0). The code is on GitHub. You can inspect it, fork it, contribute. For organizations with compliance requirements or those who want to understand what’s running on their machines, this matters.
The ReAct loop. Gemini CLI uses a Reason and Act pattern: the model explicitly reasons about what it needs to do, then acts. It’s a different pattern from Claude Code’s tool loop, but the functional result is similar.
The downsides. Gemini CLI is less mature. Code quality on complex tasks (deep refactoring, subtle dependency management) still falls below Claude Opus. Customization is more basic, no equivalent to Skills or Hooks. And the third-party ecosystem is still young.
Codex : the most open
Codex is OpenAI’s agent. Written in Rust, open source under MIT license, the most permissive of the three.
Native subagents. Codex can delegate subtasks to parallel agents. Need to refactor three independent modules? Codex spins up three subagents simultaneously. This is architecturally different from Claude Code’s multi-agent approach (Teams), more tightly integrated into the main loop.
Background streaming. Codex can work while you do something else. Launch a task, keep coding in another terminal, come back to check the result. Practical for long-running tasks.
Integrated web search and fuzzy file search with @, practical features that smooth out daily workflows.
MIT license. For companies integrating the tool into their pipelines, MIT is the simplest legally. No copyleft clause, no restrictions on commercial use.
The downsides. GPT-5-Codex code quality is inconsistent. On some tasks it’s excellent, on others it generates verbose code or misses subtleties. The plugin ecosystem is still under construction. And since the model is pay-per-use via API, costs can add up on long sessions.
Hermes Agent : the one that learns from you
Hermes Agent from NousResearch is the most original on this list. Its standout feature: a built-in learning loop. The agent creates skills from your sessions, improves them over time, and builds a persistent profile of your preferences. It’s the only one that does automatic personalization, where Claude Code requires manually maintaining a CLAUDE.md file.
Multi-provider. Hermes works with OpenRouter, Google AI Studio, OpenAI, and local models. No vendor lock-in. You can switch models mid-session.
Multi-platform. Beyond the CLI, Hermes Agent connects to Telegram, Discord, and Slack. You can interact with your coding agent from your phone.
Development pace. 4 major releases in 3 weeks (March-April 2026). The NousResearch team iterates extremely fast.
The downsides. Hermes is 3 weeks old. Stability isn’t at Claude Code’s level yet. The learning loop is promising but still v1, it can learn incorrect patterns. And the GODMODE feature (automatic model jailbreaking) raises legitimate security questions.
OpenClaw : the viral one
OpenClaw (ex-Clawdbot, ex-Moltbot) has the most chaotic trajectory of the bunch: renamed twice after an Anthropic trademark complaint, 60,000 GitHub stars in 72 hours, and its creator (Peter Steinberger) hired by OpenAI.
Persistent local memory. This is the feature that drove explosive adoption. OpenClaw remembers your sessions for weeks, no configuration needed. Where Claude Code starts fresh each conversation (except for CLAUDE.md), OpenClaw maintains continuous context.
Model agnostic. It runs with Claude, GPT, and notably Kimi 2.5, an open source model that rivals closed models on reasoning.
More than a coding agent. OpenClaw positions itself as a “Life OS” : WhatsApp connection, calendar management, third-party app control. The most ambitious approach, but also the riskiest.
The downsides. Security. 923 OpenClaw gateways were found exposed without authentication : shell access, API keys in plaintext. Massive adoption without default security is a real problem. If you deploy OpenClaw, lock it down.
Which agent for which profile?
There’s no universal answer. Here are our recommendations by use case:
Demanding solo developer : Claude Code. Reasoning quality makes the difference when you work alone and every mistake costs you time. Alternative: Gemini CLI if budget is a constraint.
Structured engineering team : Claude Code. Shared CLAUDE.md via Git, Hooks to enforce conventions, worktrees for parallel work. The investment in the harness pays off once the team exceeds 2-3 people.
Early-stage startup : Gemini CLI for daily work (free), Claude Code for critical tasks (architecture, complex refactoring). The combination optimizes the quality-to-cost ratio.
Enterprise or mid-size company : Claude Code. Governance (Skills, security Hooks, permissions) and tool maturity meet the requirements of structured teams.
Student or learning : Gemini CLI. Free, 1M context to load courses and documentation, Google Search to verify answers. Ideal for learning.
Open source contributor : Codex, Gemini CLI, or Hermes Agent. All open source, freely modifiable.
Early adopter / power user : OpenClaw. Persistent memory and multi-platform approach are unique. But secure your installation.
Multi-model experimentation : Hermes Agent. Mid-session model switching and OpenRouter support make it easy to test different LLMs on the same project.
Our pick (and why)
At Colombani.ai, we train on Claude Code. Not out of dogma, but because it produces the best code quality today, and more importantly, it offers the deepest harness : CLAUDE.md, Skills, Hooks.
But we show alternatives in every training. For a simple reason: methodology matters more than the tool. Structuring your context, planning before coding, verifying systematically, these principles work with any agent.
A well-written CLAUDE.md is a project conventions document. It works with Gemini CLI too. Skills formalize business processes, the concept transfers. The discipline of verification (tests, diff review) is universal.
The landscape is moving fast. Gemini CLI improves with every release. Codex has a solid technical architecture. Hermes Agent iterates at impressive speed. OpenClaw showed that a viral agent can emerge from the community in days. In six months, this comparison might look different. What won’t change is the value of a structured approach to AI-assisted coding.
Sources
- Claude Code : Anthropic
- Gemini CLI : Google
- Codex : OpenAI
- Hermes Agent : NousResearch
- Harness design for long-running apps : Anthropic Engineering
Going further with Colombani.ai
Turnkey offerings :
- Business AI Agent Plugins — turn a business workflow into a plugin with skills, hooks and MCP, tested and versioned. Agentic architect methodology.
- AI Security Audits — attack surface mapping, pentest, vulnerability detection, threat modeling. AI-accelerated but human-validated.
Consulting & training :
- AI maturity audit : field diagnosis and quantified roadmap in 2 weeks
- Sovereign architecture : keep your data in Europe (GDPR, AI Act)
- AI Act compliance for high-risk AI systems
- Claude Code in production : 2 days for technical teams
Freelancers and consultants who want to pass the Claude Certified Architect (CCA Foundations) certification through Colombani.ai as sponsor / employer of record : get in touch. Anthropic partnership network member.
Going deeper on Claude Code : Hooks, Skills, Claude Certified Architect.