Integration with Claude Code and/or Claude Subscription is a must

[stavarc]

For this high intensive token burner flow it is mandatory to be be able to use the Claude Subscription instead of the API

[BradKML]

BYOK across multiple services #516 (comment)

[roman-rr]

The token burn is even worse than you'd expect.

We audited ruflo's actual token consumption. While the README claims '30-50% token reduction,' it adds ~15,000-25,000 tokens of overhead per session:

• 300+ MCP tool definitions (most are non-functional stubs) loaded into context
• Intelligence hook injecting duplicate patterns every message
• Router hook printing fake ASCII tables with Math.random() latency

The '30-50% savings' claim is fabricated in source code — this.stats.totalTokensSaved += 100 hardcoded per cache hit, no actual measurement.

So you're right that subscription-level usage is needed — but much of the burn is ruflo's own overhead, not productive work. The hive-mind spawn command launches full Claude CLI processes (child_process.spawn('claude', ['--dangerously-skip-permissions', ...])), each consuming a separate API session's worth of tokens.

Alternative: Claude Code's native Agent tool does multi-agent work without the 300-tool context overhead, and it works with your existing subscription.

Full cost analysis: Independent Audit