What OpenRouter Top Apps measures and how CLI rankings differ
OpenRouter Top Apps is the platform's application leaderboard: each entry is a client that forwards prompts to OpenRouter models and accumulates weekly billed token throughput. Unlike model charts that answer "which weights get routed," Top Apps answers "which interfaces developers actually drive." That distinction matters because the same DeepSeek or Claude route can sit behind a browser IDE, a Slack bot, or a terminal agent—and those shells carry different loop depth, tool budgets, and failure modes.
For this article we apply a CLI filter: terminal-first coding agents that edit local repos, invoke shell tools, and persist sessions on a host machine. The week of June 2–8, 2026 shows CLI clients capturing a growing share of programming-related volume documented in our billing-truth rankings—teams are not merely chatting; they are running multi-hour edit-compile-test loops through OpenRouter keys. Ranking methodology is straightforward: sum tokens attributed to each app ID during the seven-day window; no self-reported surveys, no GitHub star proxies.
Star-count shopping: A CLI tops Hacker News but barely registers on OpenRouter—weekly volume exposes production adoption, not launch hype.
Mixing app classes: Hermes Agent runs multi-channel automations; comparing it to Aider without a CLI filter mis-ranks terminal-only workflows.
Ignoring host uptime: The best CLI stack still stalls when a MacBook sleeps mid-refactor—token leaders assume a machine that stays awake.
Model lock-in blindness: CLI clients default to different OpenRouter routes; picking the tool without reading its default model map duplicates spend covered in our routing guide.
Feature checklist without scenarios: MCP support alone does not tell you whether a tool fits batch migrations, iOS builds, or firefighting prod—matrix rows need task context.
Treat Top Apps as a client compass layered on top of model rankings. The platform leaderboard shows aggregate pull; the CLI slice shows what terminal engineers trust when dollars follow tokens.
Week of June 2–8, 2026: platform Top 10 and CLI-only Top 10
The first table is OpenRouter's published all-apps ranking for the window ending June 8, 2026. Totals are platform-reported token volume; positions shift weekly as clients gain or lose default-model optimizations. The second table applies the CLI filter from section 01—terminal coding agents only—so you can compare platform leaders against the subset your team deploys on a Mac or Linux shell.
| # | App | Type | Weekly tokens | Notes |
|---|---|---|---|---|
| 1 | Hermes Agent | Multi-channel agent | 4.94T | Platform #1 · broad tool orchestration · OpenRouter-native loops |
| 2 | Cursor | IDE | 2.81T | GUI-first · Agent mode drives majority of volume |
| 3 | Kilo Code | CLI | 1.22T | CLI #1 · VS Code fork path · aggressive default routing |
| 4 | Claude Code | CLI | 606B | CLI #2 · Anthropic toolchain · premium-tier bias |
| 5 | OpenClaw | Agent runtime | 541B | Persistent agents · pairs with cloud Mac setups |
| 6 | Cline | CLI / extension | 489B | Plan-and-act · strong MCP uptake |
| 7 | Roo Code | CLI / extension | 412B | Multi-mode · custom rule files |
| 8 | Aider | CLI | 378B | Git-native edits · lean dependency surface |
| 9 | Goose | CLI | 331B | Block-style recipes · enterprise OSS lineage |
| 10 | OpenCode | CLI | 297B | Terminal UI · fast iteration cycle |
Hermes Agent at 4.94T tokens widened its lead as the widest orchestration surface on OpenRouter—not a narrow repo editor, but a hub for routed sub-agents and channel connectors. Kilo Code at 1.22T (#3 platform, #1 CLI) shows terminal-first clients can outpace many GUI wrappers when default models track weekly flash leaders. Claude Code at 606B (#4 platform, #2 CLI) remains the quality-biased pick for teams already standardized on Anthropic billing.
| CLI rank | Tool | Weekly tokens | Primary strength |
|---|---|---|---|
| 1 | Kilo Code | 1.22T | High-volume edit loops · OpenRouter defaults aligned to flash MoE |
| 2 | Claude Code | 606B | Deep reasoning passes · Anthropic tool schema maturity |
| 3 | Hermes Agent | 498B CLI-attributed | Sub-agent fan-out · cross-repo automation |
| 4 | Aider | 378B | Git-diff discipline · minimal setup |
| 5 | Cline | 356B CLI-attributed | Plan mode · MCP tool breadth |
| 6 | Goose | 331B | Recipe blocks · repeatable ops playbooks |
| 7 | OpenCode | 297B | Fast TUI · low friction for small fixes |
| 8 | Codex CLI | 264B | OpenAI route integration · CI-friendly |
| 9 | Roo Code | 241B CLI-attributed | Custom modes · orchestrator/worker split |
| 10 | Qwen Code | 218B | Qwen model family optimization · cost-sensitive batches |
Platform #1 is not always CLI #1. Rank by the workflow you actually run in Terminal, then map models underneath.
CLI feature matrix: open source, MCP, sandbox, and agent modes
Token volume tells you who won last week; feature rows tell you fit. The matrix below compares the CLI Top 10 on dimensions that decide whether a tool survives enterprise review, long-running refactors, or mobile CI on a Mac host. Cells summarize June 2026 shipping behavior—verify release notes before lock-in.
| Tool | Open source | MCP | Sandbox | Sub-agents | Plan mode | Git-native | Default OR bias |
|---|---|---|---|---|---|---|---|
| Kilo Code | Yes | Yes | Partial | Limited | Yes | Yes | Flash MoE / cost |
| Claude Code | No | Yes | Yes | Yes | Yes | Yes | Anthropic premium |
| Hermes Agent | Yes | Yes | Configurable | Strong | Yes | Partial | Multi-model router |
| Aider | Yes | Via plugins | No | No | Light | Strong | User-set OR key |
| Cline | Yes | Strong | Partial | Limited | Strong | Yes | User choice |
| Goose | Yes | Yes | Recipe-level | Moderate | Yes | Partial | Enterprise config |
| OpenCode | Yes | Growing | No | No | Light | Yes | Flash defaults |
| Codex CLI | Partial | Limited | Yes | No | Moderate | Yes | OpenAI routes |
| Roo Code | Yes | Yes | Partial | Yes | Strong | Yes | Per-mode map |
| Qwen Code | Yes | Partial | No | Limited | Yes | Yes | Qwen / OR flash |
MCP separates agents that can mount databases, browsers, and ticket systems from glorified chat wrappers—Cline, Hermes, and Roo lead breadth; Aider stays intentionally minimal. Plan mode quality varies: Claude Code and Cline earn deep multi-step plans; OpenCode optimizes for quick patches. Sandbox rows matter when you run untrusted repos on a shared cloud Mac: Claude Code and Codex CLI ship stronger isolation; Aider assumes you trust the tree. Sub-agents explain Hermes Agent's platform-wide lead—fan-out work streams map cleanly to OpenRouter tiered routing from our June trends article.
No single row wins every column. High-volume teams often pair Kilo Code for daily edits with Claude Code for hard reasoning, both pointed at OpenRouter with explicit per-task model fields—mirroring the token-versus-dollar discipline in our weekly billing analysis.
Six steps to pick, configure, and run a CLI tool on a cloud Mac
Filter by CLI rank and task: Start from the CLI Top 10 table; shortlist three clients whose weekly tokens grew and whose plan/sub-agent rows match your refactor depth.
Map OpenRouter models per CLI: Document each tool's default route; override flash MoE for bulk edits and premium tiers for architecture reviews—same tiered policy as model-level routing.
Provision the host: Rent a dedicated Mac Mini M4 with SSH, static hostname, and snapshots; store API keys in Keychain, not plaintext dotfiles checked into git.
Install and pin versions: Use brew or official installers; record CLI semver in a team runbook so Monday ranking reviews coincide with controlled upgrades.
Daemonize long loops: Wrap sessions in tmux or launchd plist targets so agent runs survive disconnects; log token usage per job for finance alignment.
Validate MCP and git remotes: Test sandbox boundaries on a throwaway branch; confirm MCP servers restart after reboot—cloud Mac value is persistence, not just CPU.
Review weekly: Reconcile CLI rank shifts with your model bill; swap defaults when a client’s upstream changes OpenRouter routing. Compare tiers on the pricing page and help center.
Seven scenario recommendations (June 2026):
High-volume feature factories: Kilo Code on flash MoE defaults—matches 1.22T CLI leadership.
Architecture and security reviews: Claude Code with Anthropic tiers for plan mode depth.
Multi-repo platform ops: Hermes Agent sub-agents with tiered OpenRouter routes per worker.
Git-disciplined OSS maintenance: Aider—minimal surface, explicit diffs, user-supplied OR key.
MCP-heavy internal tools: Cline or Roo Code depending on custom mode complexity.
Runbook automation: Goose recipes for repeatable infra tasks.
Cost-sensitive batch refactors: Qwen Code or OpenCode pointed at lowest-$/T OpenRouter flash SKUs.
Teams most often skip steps 05 and 06: the first loses hours of agent work to a dropped SSH session; the second exposes production keys when MCP servers restart without Keychain access. OpenRouter ranks clients; it does not keep your host awake.
Cite-ready CLI numbers and KVMNODE cloud Mac Mini
Platform #1 (June 2–8, 2026): Hermes Agent at 4.94T weekly tokens on OpenRouter Top Apps—widest agent orchestration footprint.
CLI volume leaders: Kilo Code 1.22T (#3 platform) and Claude Code 606B (#4 platform)—terminal clients in the global Top 5.
CLI Top 10 set: Kilo Code, Claude Code, Hermes Agent, Aider, Cline, Goose, OpenCode, Codex CLI, Roo Code, Qwen Code—ten distinct terminal stacks comprising the majority of measurable CLI throughput this week.
| Runtime | CLI + OpenRouter fit | Gap | KVMNODE cloud Mac |
|---|---|---|---|
| Local MacBook | Fast to trial new CLIs | Sleep kills long agent loops | Weak for 7x24 CLI |
| Headless Linux VPS | Cheap Aider or OpenCode | No Xcode chain for iOS | Weak for Apple CI |
| Cloud Mac Mini M4 | launchd + Keychain + MCP + git | Plan rent term and snapshots | Strong for CLI agents + mobile builds |
Alternatives fail predictably: picking from stars ignores Kilo Code's 1.22T proof point; running CLI on a sleeping laptop wastes Hermes-scale routing; single-model lock-in misses the tiered switches our weekly billing guide documents. For teams that need Apple Silicon, persistent SSH, and OpenRouter-backed Kilo Code, Claude Code, or Hermes loops without a closing lid, renting a dedicated KVMNODE Mac Mini M4 / M4 Pro is the steadier path—aligned with the model-routing matrix in our OpenRouter trends post. See the pricing page and order page to deploy this week's CLI leader on hardware that stays online.