Agentic coding has become an indispensable part of modern software development. Tools like Claude Code don’t just autocomplete lines — they read entire codebases, plan and execute multi-file refactors, run tests, interpret failures, and iterate autonomously until a task is done.
Developers who adopt these workflows report dramatic reductions in time spent on boilerplate, debugging, and context-switching. For engineering teams, agentic coding is fast becoming a competitive necessity rather than a convenience.
Claude Code is Anthropic’s implementation of this paradigm. By default, it routes every prompt — along with all the code context it reads — through Anthropic’s cloud API. For many individual developers and small teams, that arrangement works fine. But for a significant portion of the industry, it is not viable. Organizations handling proprietary algorithms, patient records, financial models, source code under strict IP agreements, or government-classified data cannot send that material to an external API, regardless of the provider’s security posture. Regulated industries such as finance, healthcare, and defense operate under compliance frameworks — HIPAA, SOC 2, FedRAMP, ITAR — that explicitly restrict where data may be processed. Beyond compliance, there is the straightforward matter of cost: at scale, per-token cloud pricing accumulates quickly, and teams running continuous agentic workflows can find the economics difficult to justify. Air-gapped environments and low-latency requirements add further constraints that a remote API simply cannot satisfy.
This guide shows you how to replace the LLM backend in Claude Code with a model running entirely on your own AMD Instinct GPUs, using SGLang for serving and LiteLLM as a lightweight translation layer. The developer machine setup uses a single interactive script that handles everything from SSH verification through router deployment to the local tunnel and Claude Code launcher installation. Since every token stays on your hardware, there is no data transfer outside of your network, and no per-token billing.
By the end of this guide you will have:
-
SGLang serving GLM 5.2 on your GPU server with fp8 quantization and 8-way tensor parallelism
-
A LiteLLM router on the server translating Claude Code’s Anthropic Messages API calls into OpenAI API calls against the GLM server
-
An SSH tunnel connecting your developer machine to the router
-
Claude Code running on your developer machine — CLI or VS Code — talking to your GPU server as if it were Anthropic’s API
-
All inference running on AMD MI355X, with code and data remaining entirely on your infrastructure
Read the full guide here: Bring Claude Code On‑Prem with AMD Instinct GPUs — ROCm Blogs