If you are picking among Claude Code, Kilo Code, and Aider in mid-2026 but only read GitHub Stars, you will miss how many tokens developers burn through OpenRouter each week. This article anchors on the official OpenRouter Apps leaderboard through the week ending June 8, 2026: Hermes Agent at 4.94T, Kilo Code at 1.22T, Claude Code at 606B, plus a CLI Top 10, feature matrix, and Mac host sizing. You will leave with a comparison matrix, scenario table, and six-step checklist. Lease tiers are on the NOVAKVM pricing page; checkout on the order page.
[ SECTION_01 ] // PAIN_MAP OpenRouter Apps board vs model board: what CLI token volume actually measures
OpenRouter routes hundreds of models through one API and publishes an App usage leaderboard: integrated tools that opt into public tracking expose token consumption to everyone. Unlike the model weekly board at openrouter.ai/rankings, the Apps list measures which end product engineers invoke in production.
- Star count blind spots: Stars track attention, not daily use depth. OpenRouter tokens reflect spend times call frequency.
- Model board vs tool board: The model board shows which brain is hottest. The Apps board shows which CLI shell wraps it. DeepSeek topping models does not imply Aider tops tools.
- Sample boundary: Only apps that opt into public tracking appear. Coverage still includes Hermes Agent, Kilo Code, Claude Code, and Cline.
- Automation vs interactive: Hermes can run 24/7 in batch scripts and rack up huge token totals. That is not comparable to step-by-step Cline approval flows.
- BYOK and platform markup: Most ranked CLIs support Bring Your Own Key with no OpenRouter markup on most models.
- Host environment is underrated: Tool-call timeouts on a sleeping laptop erase whatever cost advantage the model board promised.
OpenRouter Apps rankings and the CLI Agents category page change over time. Reopen the links below before you wire production keys.
https://openrouter.ai/apps/category/coding/cli-agent
[ SECTION_02 ] // DECISION_MATRIX Week of June 2–8, 2026: platform Top 10 and CLI-focused ranking
This week (OpenRouter This Week, Monday through Sunday) shows CLI and Agent tools at roughly 70%+ of platform tokens. Hermes Agent leads at 4.94T, nearly four times OpenClaw (1.26T). Kilo Code (1.22T, #3) and Claude Code (606B, #4) sit in the Top 5.
| Rank | Tool | Type | Weekly tokens |
|---|---|---|---|
| 1 | Hermes Agent | AI Agent (CLI / personal) | 4.94T |
| 2 | OpenClaw | AI Agent (general) | 1.26T |
| 3 | Kilo Code | CLI / IDE extension | 1.22T |
| 4 | Claude Code | CLI (terminal-native) | 606B |
| 5 | Descript | AI video / podcast | 454B |
| 6–10 | pi / Lemonade / Pioneer / GitLawb / Janitor AI | Coding / vertical / chat | 218B–384B |
The CLI shortlist below filters for terminal usability and OpenRouter integration. Off-board tools show monthly totals from public tracking pages, a different window than this week.
| CLI rank | Tool | Platform this week | Open source | Core strength |
|---|---|---|---|---|
| 1 | Kilo Code | #3 · 1.22T | Yes | 500+ models, Architect / Code / Debug / Orchestrator modes |
| 2 | Claude Code | #4 · 606B | No | Terminal-native, sub-agents, macOS Seatbelt sandbox |
| 3 | Hermes Agent | #1 · 4.94T | Yes | Fully open source, general Agent, deep automation footprint |
| 4 | Aider | Off board · ~2.4B/month | Yes | Git-native, Architect dual-model cost control |
| 5 | Cline | Off board · ~140B/month | Yes | Step approval, browser automation, checkpoint rollback |
| 6 | Goose | Off board · ~46.4B/month | Yes | MCP-native, Recipes workflows, Rust build |
| 7 | OpenCode | Off board · fast growth | Yes | 75+ providers, Docker sandbox, headless CI |
| 8 | OpenAI Codex CLI | Off board · ~91B/month | Yes | Cloud sandbox, Codex-tuned fast models |
| 9 | Roo Code | Off board · ~111.8B/month | Yes | Cline fork with Boomerang sub-tasks |
| 10 | Qwen Code | Off board · ~39.9M/month | Yes | Bilingual docs, deep Qwen Coder integration |
Token volume is not who is smartest. It is which CLI gets mounted on OpenRouter and called again and again across the widest engineering surface. That is a harder selection signal than Star count.
[ SECTION_03 ] // FEATURE_MATRIX Side-by-side: open source, MCP, sandbox, Git, and model flexibility
| Feature | Kilo Code | Claude Code | Hermes Agent | Aider | Cline | Goose |
|---|---|---|---|---|---|---|
| Open source | Yes | No | Yes | Yes | Yes | Yes |
| MCP | Yes | Yes | Yes | No | Yes | Yes (native) |
| Sandbox isolation | No | System Seatbelt | No | No | Snapshot rollback | Docker |
| Sub-agent | Yes | Yes | Yes | No | Yes | Yes |
| Plan mode | Yes | Yes | Yes | No | Yes | Yes |
| Model count | 500+ | Claude only | Multi-model | 100+ | Full platform | Multi-model |
| Free BYOK | Yes | No (subscription) | Yes | Yes | Yes | Yes |
| Git integration | Yes | Strong | Yes | Strongest | Yes | Yes |
| Scenario | Recommended tool | Why |
|---|---|---|
| Daily coding with clean Git history | Aider | Git-native auto commits, Architect dual-model cost control |
| Large refactors and heavy reasoning | Claude Code | Parallel sub-agents, strong benchmarks, 606B tokens this week |
| Maximum model flexibility | Kilo Code | 500+ models, 1.22T this week, platform Top 3 |
| Security audit and step-by-step approval | Cline | Approval gates plus checkpoints, browser automation |
| DevOps and toolchain orchestration | Goose | MCP-native, Recipes for repeatable workflows |
| Tight budget and batch automation | Hermes Agent | Free open source, platform token leader this week |
| Bilingual engineering teams | Qwen Code | Mixed-language docs with deep Qwen Coder pairing |
[ SECTION_04 ] // RUNBOOK Six steps: OpenRouter key to CLI install to always-on remote Mac
- Register on OpenRouter and create keys: Visit
openrouter.ai, generate an API key on the Keys page, and issue separate keys per CLI tool so billing attribution and budget alerts stay clean. - Pick a CLI and install from official docs: Aider via
pip install aider-chat; Claude Code per Anthropic CLI instructions; Kilo Code and Cline from VS Code Marketplace or JetBrains plugin hubs; Hermes Agent from Nous Research official distribution. - Configure OpenRouter routing: Set
OPENROUTER_API_KEYin environment or config files and point Base URL tohttps://openrouter.ai/api/v1. Route batch Agent work to low-cost MoE models; reserve Claude or GPT tiers for hard reasoning. - Write project memory files: Use
CLAUDE.md,.clinerules,.goosehints, orAGENTS.mdper tool so sessions do not re-explain repo layout. - Enable plan mode and headless CI: Require plans before production edits; pin Node or Python in GitHub Actions headless modes.
- Move to a remote Mac Mini running 24/7: On dedicated Apple Silicon, pin Gateway paths, log directories, and LaunchAgent jobs. Swap models through environment variables without rebuilding hosts or losing long Agent runs to laptop sleep. SSH and always-on baselines are in the help center.
OPENROUTER_API_KEY=sk-or-v1-...
OPENROUTER_BASE_URL=https://openrouter.ai/api/v1
AIDER_MODEL=openrouter/deepseek/deepseek-v4-flash
export OPENROUTER_API_KEY OPENROUTER_BASE_URL[ SECTION_05 ] // HARDWARE Mac rental config matrix: hardware needs by CLI workload
Top CLIs bind tightly to macOS: Claude Code uses Seatbelt sandboxing; Goose runs efficiently on Apple Silicon; Kilo Code extensions pair well with Mac file permissions. Size rental hosts by intensity:
| Workload | Representative CLI | Recommended config | Notes |
|---|---|---|---|
| Light CLI | Aider, Hermes Agent | MacBook Air M2/M3 · 16GB | Most compute is cloud API; local demand stays low |
| Medium intensity | Kilo Code, Cline | MacBook Pro M3 · 16–32GB | Multi-file concurrency and browser automation need RAM |
| Heavy development | Goose plus Docker sandbox | Mac mini M4 Pro / MBP M4 Max · 32GB+ | Container sandboxes and parallel sub-agents stress memory and I/O |
| Local models | Ollama plus OpenCode | Mac Studio M4 Ultra · 64GB+ | 7B/14B local inference needs large unified memory |
For hackathons or MVP sprints, on-demand rental lets you test a higher tier before committing to purchase specs. See the pricing page.
[ SECTION_06 ] // CITABLE_FACTS Citable data snapshot and close (collected 2026-06-08, verify on official pages)
- Hermes Agent weekly tokens: 4.94T, OpenRouter platform #1; fully open source with zero subscription gate, deep penetration in automated batch pipelines.
- Kilo Code weekly tokens: 1.22T, platform #3; supports 500+ models with Architect, Code, Debug, and Orchestrator modes; BYOK with no platform markup.
- Claude Code weekly tokens: 606B, platform #4; terminal-native sub-agents with macOS Seatbelt and Linux Bubblewrap system sandboxes; Claude model family only.
- CLI and Agent share: This week, CLI and Agent tools account for roughly 70%+ of OpenRouter platform tokens. Coding workloads dominate the Apps board.
- Aider maturity signal: Public tracking shows about 2.4B tokens per month and 4.1M+ install-scale figures (pages may update). Git-native workflow and Tree-sitter repo maps drive long-term retention.
Reopen each tool official repository and OpenRouter App page before integration.
https://docs.anthropic.com/en/docs/claude-code
June 2026 shows a clear CLI split: Hermes Agent leads on scaled open automation; Kilo Code and Claude Code own multi-model flexibility vs strongest reasoning; off-board Aider, Cline, and Goose still win on Git workflow and audit gates.
Chasing the board on sleeping laptops or jittery SSH means Kilo Code 1.22T weekly tokens never becomes merged PRs. Gateway drops, disk exhaustion, Docker OOM, and MCP timeouts will not appear on OpenRouter charts, yet they cap real success rates.
For Claude Code remote Gateway, Kilo Code orchestration, or Hermes Agent 24/7 automation, pair weekly routing reviews with a dedicated Apple Silicon bare-metal host. NOVAKVM offers multi-region Mac Mini M4 and M4 Pro elastic leases. See the pricing page and order page.