vLLM powers local Claude Code swarm

Local agent stacks are starting to look like real infrastructure instead of a hobbyist stunt. vLLM is the key enabler here: it gives Claude Code a localhost backend, parallel inference, and the kind of batching behavior that makes multi-agent work feel practical.

• –vLLM's OpenAI-compatible server and Docker path make it easy to plug into existing agent tooling without rewriting workflows.
• –OpenAI's gpt-oss-120b is built for efficient local inference and tool use, which fits the offline, self-hosted setup.
• –The bottleneck shifts from "can I run this?" to "how many agents can my GPU keep fed?" which is the right problem to have.
• –This is a strong argument for Linux plus local inference if you care about privacy, latency, or avoiding per-request cloud costs.