First-week June token data · CLI top 10 · Kilo Code / Claude Code / Hermes Agent · feature matrix · Mac rental configs
If you are choosing an AI command-line assistant on Mac or Windows, GitHub stars and launch demos tell an incomplete story. Who is actually being called—and how many tokens they burn each week—is closer to production reality. Bottom line: using OpenRouter Top Apps data for June 2–8, 2026, Hermes Agent leads the full platform at 4.94T tokens, while Kilo Code (1.22T) and Claude Code (606B) sit in the global top five as the coding CLI duo. This guide covers: Apps leaderboard methodology, platform top ten, CLI-specific top ten, deep dives, a feature matrix, scenario-based picks, Mac rental hardware guidance, and FAQ. Pair with our LLM trends and weekly model rankings articles.
OpenRouter is a unified LLM routing platform: one API key, hundreds of models (Claude, GPT, Gemini, DeepSeek, Llama, and more), BYOK with no platform markup, failover routing, and a public App usage leaderboard. Tools that opt into public tracking expose real token consumption—one of the most objective heat maps in AI tooling today.
Stars measure attention; tokens measure workload. This article filters OpenRouter’s platform data down to CLI and Agent developer tools, then blends GitHub stars, open-source status, developer experience, and feature depth into a CLI-specific ranking.
Stars skew toward curiosity: bookmarking does not equal daily active use.
Tokens skew toward habit: weekly rolls show who keeps routing traffic.
Sample caveat: only publicly tracked apps appear—not exhaustive, but covers mainstream CLIs.
Snapshot: June 2–8, 2026 (OpenRouter “This Week”). Source: openrouter.ai/apps. Rankings move weekly—verify live data when you read this.
| Rank | App | Category | Weekly tokens |
|---|---|---|---|
| 1 | Hermes Agent | AI Agent (CLI/personal) | 4.94T |
| 2 | OpenClaw | General AI Agent | 1.26T |
| 3 | Kilo Code | CLI / IDE extension | 1.22T |
| 4 | Claude Code | Terminal-native CLI | 606B |
| 5 | Descript | AI video/podcast editor | 454B |
| 6 | pi | Coding AI | 384B |
| 7 | Lemonade | Roblox AI tool | 317B |
| 8 | Pioneer (production) | Inference API | 268B |
| 9 | GitLawb | Blockchain collab | 253B |
| 10 | Janitor AI | AI chat platform | 218B |
Trend line: CLI and Agent tools together account for roughly 70%+ of this week’s tokens. Kilo Code and Claude Code are the coding CLI standouts. Non-dev apps (Descript, Lemonade) prove OpenRouter’s audience has expanded beyond pure developers.
Blending this week’s live data, trailing 30-day volume, and feature scoring (entertainment-only apps excluded):
| CLI rank | Tool | Platform week | Tokens | Stars | Open | Highlight |
|---|---|---|---|---|---|---|
| 1 | Kilo Code | #3 | 1.22T | 16,200+ | Yes | 500+ models, four modes |
| 2 | Claude Code | #4 | 606B | N/A | No | Top reasoning, sub-agents |
| 3 | Hermes Agent | #1 | 4.94T | Active | Yes | Fully open, massive reach |
| 4 | Aider | Off chart | ~2.4B/mo | 41,200+ | Yes | Git-native maturity |
| 5 | Cline | Off chart | ~140B/mo | 58,600+ | Yes | Approve-every-step UX |
| 6 | Goose | Off chart | ~46.4B/mo | 32,300+ | Yes | MCP-native workflows |
| 7 | OpenCode | Off chart | Fast growth | 97,500+ | Yes | 75+ providers |
| 8 | OpenAI Codex CLI | Off chart | ~91B/mo | 62,000+ | Yes | Cloud sandbox |
| 9 | Roo Code | Off chart | ~111.8B/mo | Active | Yes | Cline fork, task boards |
| 10 | Qwen Code | Off chart | ~39.9M/mo | Active | Yes | Chinese-friendly |
At 1.22T tokens this week—only 40B behind OpenClaw—Kilo Code shows deep daily use. It supports 500+ models, four modes (Architect, Code, Debug, Orchestrator), zero-markup BYOK, VS Code and JetBrains, plus MCP integrations. Best for teams that swap models often and orchestrate multi-step work inside the IDE.
Anthropic’s terminal-native coding agent hit 606B tokens (platform #4). Strengths: parallel sub-agents, MCP, CLAUDE.md memory, macOS Seatbelt / Linux Bubblewrap sandboxes, Plan Mode, headless CI. Trade-offs: Claude-only, higher cost at scale, closed source. Ideal for large refactors with budget headroom.
Nous Research’s fully open agent burned 4.94T tokens—nearly 4× the runner-up. Free, no subscription, general-purpose (not coding-only), deep OpenRouter integration. Volume reflects heavy automation deployments. Strong fit for researchers, students, and teams that need self-hosted control.
Aider (#4): most mature Git-native CLI; architect dual-model cost control; Tree-sitter repo maps save tokens; 41k+ stars; no MCP or sub-agents. Cline (#5): approve-everything philosophy; headless browser automation; workspace checkpoints—great for audit-sensitive teams. Goose (#6): from Block; MCP-first with 1,700+ servers; Recipes for reusable flows; Rust build shines on Apple Silicon.
OpenCode (#7): fastest star growth (97.5k+); Docker sandbox; 650k+ active users. Codex CLI (#8): OpenAI cloud sandbox; ~1000 tokens/sec. Roo Code (#9): enhanced Cline fork with Boomerang task splitting. Qwen Code (#10): bilingual Chinese/English optimization tied to Qwen2.5-Coder.
| Feature | Kilo | Claude | Hermes | Aider | Cline | Goose | OpenCode |
|---|---|---|---|---|---|---|---|
| Open source | Yes | No | Yes | Yes | Yes | Yes | Yes |
| MCP | Yes | Yes | Yes | No | Yes | Yes | Yes |
| Sandbox | No | OS-level | No | No | Snapshot | Docker | Docker |
| Sub-agents | Yes | Yes | Yes | No | Yes | Yes | Yes |
| Free BYOK | Yes | No | Yes | Yes | Yes | Yes | Yes |
| Git integration | Yes | Yes | Yes | Best | Yes | Yes | Yes |
| Browser automation | No | No | No | No | Yes | No | No |
| Scenario | Pick | Why |
|---|---|---|
| Clean Git history daily | Aider | Native commits, dual-model savings |
| Large refactor, budget OK | Claude Code | Best reasoning, parallel agents |
| Maximum model choice | Kilo Code | 500+ models, platform top 3 |
| Audit every action | Cline | Per-step approval, rollback |
| DevOps automation | Goose | MCP recipes, 1,700+ services |
| Free personal use | Hermes Agent | Open, zero subscription |
| Chinese dev teams | Qwen Code | Bilingual docs and comments |
Register at openrouter.ai and create an API key.
Export OPENROUTER_API_KEY in your shell profile.
Run a 30-minute proof of concept with one or two tools from the matrix above.
Re-check Apps rankings weekly to validate routing choices.
On macOS, complete sandbox, Keychain, and MCP permissions in a graphical session (see section 08).
Top CLIs bind tightly to macOS: Claude Code uses Seatbelt sandboxes; Goose (Rust) runs efficiently on M-series chips; Kilo Code pairs cleanly with macOS file permissions in VS Code; Aider leans on Homebrew and pyenv ecosystems that are strongest on Mac. Unified memory makes Apple Silicon the de facto platform for AI coding assistants.
| Workload | Recommended Mac | Notes |
|---|---|---|
| Light CLI (Aider, Hermes) | MacBook Air M2/M3, 16GB | Cloud API–heavy; low local compute |
| Mid (Kilo Code, Cline) | MacBook Pro M3, 16–32GB | Multi-file edits, browser automation |
| Heavy (Goose + Docker) | Mac mini M4 Pro, 32GB+ | Parallel agents, container I/O |
| Local models (Ollama + OpenCode) | Mac Studio M4 Ultra, 64GB+ | 7B–14B on-device inference |
Rental upside: hackathons, short contracts, team onboarding, and hardware trials before a five-figure Studio purchase. Even cloud-only CLIs benefit from Mac efficiency—quieter fans and longer battery life matter when agents run for hours.
Calendar week Monday–Sunday. This article uses June 2–8, 2026. Always verify at openrouter.ai/apps.
Most models have no markup. BYOK pricing tracks direct API rates. Some failover routes may add a small surcharge.
Claude Code (OS sandboxes) and Cline (per-step approval plus snapshots) lead. Goose and OpenCode add Docker isolation for execution.
Free, open, and widely scripted in unattended automation. One pipeline can dwarf interactive CLI sessions.
The first week of June 2026 shows a split CLI landscape: Hermes Agent owns raw volume, Kilo Code anchors developer preference, Claude Code holds the reasoning crown, and Aider plus Cline remain irreplaceable for Git hygiene and compliance. Do not pick from stars alone—cross-check token thermometers with your scenario matrix.
Without a Mac—or with Windows as your daily driver—hidden costs show up in sandbox prompts, Keychain approvals, lid-close agent drops, and OAuth flows SSH cannot click. You can choose the right tool from this leaderboard and still fail Gateway setup on the wrong OS. Seatbelt, browser automation, and Docker sandboxes all benefit from a real macOS graphical session for first-run acceptance.
To run Kilo Code, Claude Code, or Hermes Agent against OpenRouter Apps data on macOS long term, rent a physical Mac node through VNCMac: use the primary button below for pricing, and map hardware to section 08 before you commit to a purchase.