OPENROUTER CLI
TOP_APPS_
MAC_AGENT_
PICK_2026.
GitHub Stars lie; settled OpenRouter Top Apps token volume does not. Pain point: Mac developers bounce between Hermes Agent, Kilo Code, Claude Code, Aider, and Cline without separating model rankings from application rankings — the former tells you which LLM to route; the latter shows which agent runtime is actually burned in production. Conclusion: week of June 2–8, 2026, Hermes Agent 4.94T leads the platform, Kilo Code 1.22T and Claude Code 606B break into the global Top 5; CLI+Agent tools consume 70%+ of weekly tokens. Roadmap: how to read the apps chart → platform snapshot → CLI Top 10 → feature matrix → scenario picks → five deployment steps → case study → Mac rental matrix.
1. Pain Points: Apps Chart vs Model Chart
(1) Model rankings answer "which LLM": OpenRouter rankings Top Models and Programming Collections show whether DeepSeek-V4-Flash or Hy3 preview leads — that belongs in openclaw.json or Cursor fallback chains. (2) App rankings answer "which tool": openrouter.ai/apps tracks clients that opt into public telemetry — Hermes, Kilo, Claude Code token share reflects runtime + workflow penetration. (3) Stars ≠ throughput: OpenCode ships 97,500+ GitHub stars but sits off this week's apps leaderboard; Aider's 41,200+ stars mask lower interactive token burn versus batch automation agents. (4) Mac coupling: Claude Code's Seatbelt sandbox, Goose's Rust perf on Apple Silicon, Kilo's VS Code integration on macOS file permissions — top CLI tools bind tightly to Mac; hardware must be part of the decision.
2. Data Window: June 2–8, 2026
Source: OpenRouter Top Apps, "This Week" slice (Mon–Sun), snapshot date June 8, 2026. Only publicly tracked apps; not a census but covers mainstream CLI/IDE agents. CLI-specific ordering blends weekly platform data, 30-day cumulative usage, terminal UX, MCP/sandbox support, OSS status, and community velocity. Non-dev entertainment apps (e.g. Janitor AI) excluded from the CLI lens.
3. Platform Top 10: CLI Duo in Top 5
| Rank | App | Type | Weekly Tokens | Note |
|---|---|---|---|---|
| 1 | Hermes Agent | CLI Agent | 4.94T | Platform #1, OSS zero-friction |
| 2 | OpenClaw | General Agent | 1.26T | IM automation, popular on Mac |
| 3 | Kilo Code | CLI / IDE | 1.22T | 500+ models, four modes |
| 4 | Claude Code | Terminal CLI | 606B | Closed-source reasoning ceiling |
| 5 | Descript | AV editor | 454B | Non-dev; platform breadth signal |
CLI and Agent-class tools together exceed 70% of weekly tokens. Hermes at 4.94T is nearly 4× OpenClaw (1.26T) — batch automation, research pipelines, and headless agents dominate token curves versus IDE chat loops.
4. CLI-Specific Top 10
| CLI # | Tool | Platform Week | Tokens | OSS | Highlight |
|---|---|---|---|---|---|
| 1 | Kilo Code | #3 | 1.22T | Yes | 500+ models, Architect/Code/Debug/Orchestrator |
| 2 | Claude Code | #4 | 606B | No | Sub-agents, Seatbelt, Plan Mode |
| 3 | Hermes Agent | #1 | 4.94T | Yes | Fully OSS, cross-session memory |
| 4 | Aider | Off chart | ~2.4B/mo | Yes | Git-native, Tree-sitter repo map |
| 5 | Cline | Off chart | ~140B/mo | Yes | Step approval, browser automation |
| 6 | Goose | Off chart | ~46.4B/mo | Yes | MCP-native, 1700+ integrations |
| 7 | OpenCode | Off chart | Rapid growth | Yes | 75+ providers, Docker sandbox |
| 8 | OpenAI Codex CLI | Off chart | ~91B/mo | Yes | Cloud sandbox, Codex model speed |
| 9 | Roo Code | Off chart | ~111.8B/mo | Yes | Cline fork, Boomerang tasks |
| 10 | Qwen Code | Off chart | ~39.9M/mo | Yes | Bilingual CN/EN, Qwen2.5-Coder |
5. Feature Matrix
| Capability | Kilo | Claude Code | Hermes | Aider | Cline | Goose | OpenCode |
|---|---|---|---|---|---|---|---|
| MCP | Yes | Yes | Yes | No | Yes | Yes++ | Yes |
| Sandbox | No | Seatbelt | No | No | Checkpoint | Docker | Docker |
| Sub-agent | Yes | Yes | Yes | No | Yes | Yes | Yes |
| Free BYOK | Yes | No | Yes | Yes | Yes | Yes | Yes |
| Model count | 500+ | Claude only | Multi | 100+ | All | Multi | 75+ |
6. Seven Scenario Picks
A · Clean Git history → Aider. B · Large refactor + budget → Claude Code (~4% of GitHub AI-assisted commits). C · Model flexibility → Kilo Code (1.22T proves depth). D · Security audit trail → Cline step approval. E · DevOps orchestration → Goose Recipes + MCP. F · Tight budget → Hermes Agent (free OSS) or watch Gemini CLI policy shifts. G · Chinese dev → Qwen Code.
7. Five Deployment Steps on Mac
Step 1 — Monday snapshot at openrouter.ai/apps/cli-agent
Archive rank deltas; >20% week-over-week on Hermes/Kilo triggers agent residency review.
Step 2 — Split runtime vs model routing
Pick Hermes/OpenClaw/Kilo as runtime; model layer still follows weekly Programming Collections — never merge both into one config file.
Step 3 — Tag Mac three tiers: local / OpenRouter API / remote Mac
Light Aider/Hermes on 16GB Air; Kilo+Cline browser automation on M3 Pro 32GB; Goose+Docker on remote Mac mini M4 Pro 32GB+.
Step 4 — Cap agent spend; isolate Headless CI
Claude Code Pro from $20/mo plus usage; BYOK stacks hard-cap OpenRouter at e.g. $500/mo with auto-fallback to DeepSeek-V4-Flash.
Step 5 — 20-task bake-off across three stacks
Same refactor on Hermes, Kilo, Claude Code — measure time, $/task, rollback count. Beats reading rankings alone.
8. Mac Rental Configuration Matrix
| Workload | Hardware | Stack | Rationale |
|---|---|---|---|
| Light CLI | MacBook Air M2/M3 16GB | Aider, Hermes | API-bound; minimal local compute |
| Medium | MacBook Pro M3 16–32GB | Kilo, Cline | Multi-file + browser automation RAM |
| Heavy sandbox | Mac mini M4 Pro 32GB+ | Goose, OpenCode Docker | Parallel agents + container I/O |
| Local fallback | Mac Studio M4 Ultra 64GB+ | Ollama + OpenCode | 7B/14B unified memory floor |
9. Case Study: Why CLI Tools Eat 70%+ Tokens
"An 8-person remote team ran Goose + Cline on Windows/WSL; OpenRouter billed $4,100/mo and Docker sandbox I/O jittered on NTFS mounts. After mapping June week-1 apps data: (1) gateway automation moved to Hermes; interactive work to Kilo Code; (2) 58% model traffic to DeepSeek-V4-Flash; (3) everyone shifted agent sandboxes to MACGPU M4 Pro 32GB remote nodes. Four weeks later agent tokens down 31%, P95 sandbox boot 42s → 11s. Windows laptops kept Cursor review only — proof that stack pick + hardware tiering beats model hopping alone."
Industry read: the apps chart signals agentized development at scale. Hermes 4.94T reflects mass deployment in automation and research; Kilo 1.22T shows multi-model IDE agents are production-default; Claude Code 606B carries higher ARPU — quality track and volume track coexist. Apple Silicon unified memory enables local Ollama fallback beside cloud agents; macOS Seatbelt makes Claude Code's security story testable on real hardware — why Mac is the de facto AI dev platform, not marketing folklore.
10. Hard Numbers & Acceptance Checklist
Hermes weekly 4.94T (#1). Kilo 1.22T (#3). Claude Code 606B (#4). CLI+Agent share 70%+. Aider installs 4.1M+; Cline stars 58,600+. Claude Code ~4% of GitHub AI-assisted commits.
Checklist: CLI apps chart screenshot □ | Runtime vs model configs split □ | Mac three-tier tags □ | BYOK monthly cap □ | 20-task bake-off □ | Docker node ≥32GB □ | Remote Mac launchd residency □
Windows and Linux run these CLIs fine, but parallel Xcode/FCP/ComfyUI, Claude Code Seatbelt, launchd 24/7 Hermes/OpenClaw, Metal-sidecar MLX validation integrate cleaner on macOS. If 16GB local RAM is saturated by Docker sandboxes or thermals throttle overnight agents, MACGPU remote Mac nodes (M3 Pro 32GB / Mac mini M4 Pro) can host Goose/Hermes headless while your laptop runs Kilo/Cursor review — rental compute buys predictable monthly burn and stable throughput.