← Hacker News

Show HN: Dragoman – Multi-model routing for Claude Code via sub-agents

Hacker News · asakin · May 12, 2026
Dragoman is a CLI tool that integrates with Claude Code's sub-agent system to route queries to different AI models (Perplexity, Gemini, Ollama) based on context, allowing users to leverage the most appropriate model without switching applications. The tool synthesizes responses from multiple models and securely resolves API keys from 1Password or Keychain at runtime without exposing them to Claude's context.

Detailed Analysis

Dragoman, an approximately 800-line command-line interface tool, addresses a practical friction point for power users of Claude Code who simultaneously maintain subscriptions to multiple AI services: the need to manually switch between tools depending on which model is best suited for a given task. Built by a developer running Claude Code alongside Perplexity, OpenAI, Gemini, and a local Ollama instance, the tool inserts itself into Claude Code's existing sub-agent architecture to perform intent-based routing — directing news queries to Perplexity, explicit model requests to their respective targets, and computationally sensitive queries to a local model — without requiring the user to leave the Claude Code environment.

The technical approach is notable for what it does not build. Rather than constructing a parallel orchestration layer, Dragoman leverages Anthropic's own sub-agent system as its execution backbone, with the tool itself described as simply "adding the verb." This design philosophy reflects an increasingly common pattern in the AI tooling ecosystem: building thin, composable layers atop existing model infrastructure rather than attempting to replicate or replace it. The security model is similarly minimal but deliberate — API keys for third-party services are resolved at call time from system credential stores such as 1Password or macOS Keychain, ensuring they never enter Claude's active context window, a meaningful consideration given that context contents can in principle be exposed through prompt injection or logging.

The fan-out capability — the ability to simultaneously query up to four models and return a Claude-synthesized result — points to a broader pattern of ensemble reasoning that has grown more feasible as per-token costs decline and latency improves across frontier models. Rather than treating model selection as a binary choice, Dragoman treats it as a portfolio decision, acknowledging that different models carry different strengths across retrieval, reasoning, code generation, and real-time information access. The synthesis step, handled by Claude, effectively positions Anthropic's model as a meta-reasoner over the outputs of its competitors.

At a structural level, this project reflects the maturing of Claude Code's extensibility surface as a legitimate developer platform. Anthropic's sub-agent architecture, originally designed to allow Claude to delegate discrete tasks to specialized agents within a workflow, is here being repurposed as a routing bus for heterogeneous model backends. The fact that this was achievable in roughly 800 lines of code suggests the abstraction Anthropic has provided is both well-scoped and genuinely composable. For Anthropic, this kind of third-party tooling reinforces Claude Code's position as a workflow hub rather than merely an assistant, even as it channels competitive model traffic through Claude's orchestration layer.

The broader trend this project exemplifies is the normalization of multi-model workflows among technically sophisticated users who view no single model as universally optimal. As AI services proliferate and subscription fatigue grows, tooling that reduces context-switching costs while preserving model diversity will likely see increasing demand. Dragoman represents an early, lightweight instantiation of what may become a standard component in developer AI workflows: a lightweight routing layer that treats model selection as a dynamic, task-dependent decision rather than a static configuration choice.

Read original article →