If you still pick an AI coding assistant by GitHub Stars while ignoring real token burn on OpenRouter, the App usage leaderboard for the first week of June 2026 (Jun 2–8) tells a sharper story: Hermes Agent leads the entire platform at 4.94T tokens, with Kilo Code (1.22T) and Claude Code (606B) in the Top 5. This guide is for developers choosing an AI command-line tool. It covers the CLI-focused Top 10, a 13-dimension feature matrix, seven scenario picks, plus a six-step rollout checklist and Mac rental hardware tiers.
OpenRouter is one of the largest neutral AI model API aggregators: one API key reaches 300+ models, and the platform publishes a live App usage leaderboard. When a connected tool opts into public tracking, its token volume and request count are visible to everyone—one of the most objective signals of AI tool adoption today.
Stars measure attention; tokens measure depth: Cline has 58,600+ Stars, but the OpenRouter weekly board reflects how many Agent loops run every day.
CLI and Agent tools account for 70%+ of this week's tokens: coding automation has moved from IDE plugins to terminal-native Agents—the top of the chart is almost all CLI / Agent tools.
BYOK removes platform markup: most CLI tools accept your own API key, so the leaderboard reflects real model calls—not marketing spend.
Usage is not the same as quality: high token counts can come from batch automation (Hermes), not stronger single-shot reasoning—match tools to your scenario.
Data is time-sensitive: the "This Week" view shifts every week; check openrouter.ai/apps for the latest numbers.
Pair with the model leaderboard: model rankings answer "which brain"; App rankings answer "which hands"—read both together; see the OpenRouter weekly token rankings guide.
"It is not who has the most Stars—it is who burns the most tokens. The App usage board is a thermometer for real CLI tool adoption."
As of June 8, 2026, here is the OpenRouter "This Week" App token consumption ranking. Among CLI tools, Kilo Code (#3) and Claude Code (#4) stand out.
| Platform rank | Tool | Type | Weekly tokens | CLI-related |
|---|---|---|---|---|
| 1 | Hermes Agent | AI Agent (CLI / personal) | 4.94T | Yes — open general Agent |
| 2 | OpenClaw | AI Agent (general) | 1.26T | Yes — Gateway + CLI |
| 3 | Kilo Code | CLI / IDE extension | 1.22T | Yes — 500+ models |
| 4 | Claude Code | CLI (terminal-native) | 606B | Yes — Anthropic official |
| 5 | Descript | AI video / podcast | 454B | — |
| 6 | pi | Coding AI | 384B | Partial |
| 7–10 | Lemonade / Pioneer / GitLawb / Janitor AI | Non-dev | 218–317B | — |
| CLI rank | Tool | Weekly tokens | GitHub Stars | Open source | Core strength |
|---|---|---|---|---|---|
| #1 | Kilo Code | 1.22T (platform #3) | 16,200+ | Yes | 500+ models, four work modes |
| #2 | Claude Code | 606B (platform #4) | N/A | No | Strongest reasoning, Sub-agent orchestration |
| #3 | Hermes Agent | 4.94T (platform #1) | Active | Yes | Fully open source, high adoption |
| 4 | Aider | ~2.4B/month | 41,200+ | Yes | Git-native, most mature CLI |
| 5 | Cline | ~140B/month | 58,600+ | Yes | Human-in-the-loop, step approval |
| 6 | Goose | ~46.4B/month | 32,300+ | Yes | MCP-native, workflow orchestration |
| 7 | OpenCode | Rapid growth | 97,500+ | Yes | 75+ models, fastest growth |
| 8 | OpenAI Codex CLI | ~91B/month | 62,000+ | Yes | Cloud sandbox, fast response |
| 9 | Roo Code | ~111.8B/month | Active | Yes | Highly customizable, task tracking |
| 10 | Qwen Code | ~39.9M/month | Active | Yes | Alibaba, Chinese-friendly |
Citable hard data: (1) Hermes Agent hit 4.94T tokens this week—nearly 4x second-place OpenClaw (1.26T). (2) Kilo Code at 1.22T trails OpenClaw by only 40B, signaling very high daily active usage. (3) Claude Code at 606B continues to lead on SWE-bench and similar benchmarks. (4) Aider has 4.1M+ installs and processes roughly 15 billion tokens per week.
Methodology: The CLI-focused list filters tool types from platform data, weighting 30-day totals and feature completeness. Non-dev / entertainment apps are excluded. Snapshot date: 2026-06-08; confirm against the live leaderboard.
Rankings tell you what is hot; a feature matrix tells you what fits your team. This table covers the 13 dimensions that matter most for CLI selection.
| Feature | Kilo Code | Claude Code | Hermes | Aider | Cline | Goose | OpenCode |
|---|---|---|---|---|---|---|---|
| Open source | Yes | No | Yes | Yes | Yes | Yes | Yes |
| MCP support | Yes | Yes | Yes | No | Yes | Yes | Yes |
| Sandbox isolation | No | Yes — system-level | No | No | Snapshot | Docker | Docker |
| Sub-agent | Yes | Yes | Yes | No | Yes | Yes | Yes |
| Plan mode | Yes | Yes | Yes | No | Yes | Yes | Yes |
| Model count | 500+ | Claude only | Multi-model | 100+ | All platforms | Multi-model | 75+ |
| Free BYOK | Yes | No | Yes | Yes | Yes | Yes | Yes |
| Git integration | Yes | Yes | Yes | Yes — strongest | Yes | Yes | Yes |
| Browser automation | No | No | No | No | Yes | No | No |
Kilo Code: VS Code / JetBrains extension plus CLI. Four modes—Architect, Code, Debug, Orchestrator—with 500+ models and zero-markup BYOK. Its 1.22T weekly tokens signal deep daily usage.
Claude Code: Anthropic's terminal-native Agent with Sub-agent parallelism, MCP, CLAUDE.md project memory, and macOS Seatbelt sandboxing. Claude-only models from $20/month, but reasoning leads on SWE-bench.
Hermes Agent: Nous Research's fully open Agent, leading at 4.94T tokens this week. Free, self-hostable, with Telegram Gateway and scheduled tasks—ideal for batch automation and 24/7 always-on workloads (see the Hermes Agent Mac rental deploy guide).
# Configure OpenRouter in any CLI tool (environment variables) export OPENROUTER_API_KEY="sk-or-v1-xxxxxxxx" export OPENROUTER_BASE_URL="https://openrouter.ai/api/v1" # Aider example aider --model openrouter/anthropic/claude-sonnet-4 # Verify your key curl -s https://openrouter.ai/api/v1/auth/key \ -H "Authorization: Bearer $OPENROUTER_API_KEY"
Note: Hermes Agent's high token volume partly reflects large-scale batch deployments. Do not compare that directly to interactive tools like Aider or Claude Code on single-shot reasoning quality.
There is no universal best CLI—only the best match for your workflow. Cross-check this table against the feature matrix above.
| Scenario | Recommended tool | Why | Team size |
|---|---|---|---|
| A: Daily coding + clean Git history | Aider | Git-native workflow, dual-model cost control | Individual / small team |
| B: Large refactors + budget headroom | Claude Code | Strongest reasoning, Sub-agent parallelism | Mid / large team |
| C: Maximum model flexibility | Kilo Code | 500+ models, four work modes | Small team (2–10) |
| D: Security-sensitive / step audit | Cline | Per-step approval + workspace snapshot rollback | Compliance teams |
| E: DevOps / workflow automation | Goose | MCP-native 1700+ services, Recipes system | Platform engineering |
| F: Tight budget / free tier | Hermes Agent | Fully free and open source, #1 weekly tokens | Individual developers |
| G: Chinese developers / Alibaba ecosystem | Qwen Code | Bilingual optimization, Qwen2.5-Coder integration | Chinese-speaking community |
Top-ranked tools have deep ties to macOS: Claude Code's sandbox uses macOS Seatbelt; Goose is Rust-built with strong Apple Silicon performance; Aider's Python stack has the richest toolchain on macOS (Homebrew, pyenv). Mac is the de facto standard platform for AI coding tools.
Choosing a tool is step one. Pinning your CLI Agent to a predictable execution environment avoids hidden costs like laptop sleep killing long sessions or cheap VPS hosts that cannot run xcodebuild.
Check openrouter.ai/apps every Monday: log CLI App rank changes for an internal weekly report; cross-reference the model trends guide.
Pick 1–2 primary CLIs by scenario: avoid five tools per developer—Aider (Git flow) + Claude Code (refactors) or Kilo Code (multi-model) covers about 80% of cases.
Configure OpenRouter BYOK: set OPENROUTER_API_KEY; use cheap Flash models for batch work and Sonnet/Opus for complex reasoning to control monthly spend.
Check whether local Mac memory is enough: multi-Agent runs + Docker sandboxes (Goose/OpenCode) + browser automation (Cline) can swap heavily on a 16GB MacBook Air.
Match Mac tier to tool (see table below): lightweight CLIs suit MacBook Air M2/M3 16GB; heavy dev needs Mac mini M4 Pro 32GB+; local Ollama needs Mac Studio 64GB+.
SSH into a dedicated cloud Mac: long CLI Agent sessions, 24/7 Hermes Gateway, and xcodebuild pipelines belong on an SSH node; APIs handle elastic peaks. See rental rates.
| Use case | Recommended Mac | Notes |
|---|---|---|
| Light CLI (Aider, Hermes) | MacBook Air M2/M3, 16GB | Mostly cloud API; low local compute need |
| Medium (Kilo Code, Cline) | MacBook Pro M3, 16–32GB | Multi-file concurrency and browser automation need RAM |
| Heavy (Goose + Docker) | Mac mini M4 Pro, 32GB+ | Docker sandbox + parallel Agents stress memory and I/O |
| Local models (Ollama + OpenCode) | Mac Studio M4 Ultra, 64GB+ | Running 7B/14B models locally needs large unified memory |
A sleeping laptop, shared Linux VPS, or headless-only host cannot sustain 12-hour Agent loops or run macOS-only tools like notarytool and xcodebuild. Pairing "read the CLI rankings" with "fix the execution environment" is more sustainable than chasing the hottest tool every week.
Teams that need stable SSH long sessions, Keychain isolation, and predictable bandwidth for iOS CI/CD and AI Agent automation should place heavy workloads on a dedicated cloud Mac rather than betting everything on a local laptop or cheap VPS. NodeMini Mac Mini cloud rental fits the CLI Agent execution layer: swap API keys or tools while SSH nodes and CI labels stay fixed. See the help center for access and compute ordering for instant provisioning.
GitHub Stars reflect attention and community size. The OpenRouter App board sorts by real token consumption, which better reflects daily call frequency and usage depth. Combine both, but weight usage data higher. See hardware tiers in rental rates.
Hermes Agent is fully free and open source, widely deployed in automation scripts, research, and batch pipelines—a single workflow can generate massive token volume. Claude Code is mostly interactive sessions with more controlled per-session usage, but leads benchmarks on single-shot reasoning quality.
Lightweight CLIs (Aider, Hermes) run on MacBook Air M2/M3 16GB. Docker sandboxes or multi-Agent parallelism need Mac mini M4 Pro 32GB+. Local Ollama inference needs Mac Studio 64GB+ unified memory. Project-based rentals let you switch tiers by scenario—see the help center and the SSH session isolation guide.