nvd-claude-proxy 0.2.8

Run Claude Code (and any Anthropic SDK client) on NVIDIA NIM models via a local proxy.

nvd-claude-proxy

Run Claude Code — and any Anthropic SDK client — on free NVIDIA NIM models.

A local HTTP proxy that speaks the Anthropic Messages API and forwards requests to https://integrate.api.nvidia.com/v1 (NVIDIA NIM / build.nvidia.com). Point your ANTHROPIC_BASE_URL at the proxy and your tools work unchanged while the inference runs on Nemotron Ultra, Qwen3, DeepSeek-R1, or any other NIM model.

---

Install

pip install nvd-claude-proxy

# Optional extras
pip install "nvd-claude-proxy[metrics]"   # Prometheus /metrics endpoint
pip install "nvd-claude-proxy[pdf]"       # PDF document block extraction
pip install "nvd-claude-proxy[full]"      # everything above

Quick start

# 1. Get a free API key at https://build.nvidia.com  (no credit card required)
export NVIDIA_API_KEY=nvapi-...

# 2. Start the proxy (default port 8788)
nvd-claude-proxy

In another shell (or in your shell profile):

export ANTHROPIC_BASE_URL=http://localhost:8788
export ANTHROPIC_API_KEY=not-used          # any non-empty string works
export ANTHROPIC_MODEL=claude-opus-4-7     # → Nemotron Ultra 253B
export ANTHROPIC_SMALL_FAST_MODEL=claude-haiku-4-5  # → Nemotron Nano 9B v2
claude                                     # launch Claude Code

No config file required — the default models.yaml is bundled in the package.

---

Default model mapping

Claude alias	NVIDIA NIM model	Notes
claude-opus-4-7	nvidia/llama-3.1-nemotron-ultra-253b-v1	Reasoning, best quality
claude-sonnet-4-6	nvidia/llama-3.3-nemotron-super-49b-v1	Balanced
claude-haiku-4-5	nvidia/nvidia-nemotron-nano-9b-v2	Fast, small
claude-opus-4-7-vision	meta/llama-4-maverick-17b-128e-instruct	Vision-capable
claude-qwen3	qwen/qwen3-235b-a22b	Qwen3 thinking
claude-r1	deepseek-ai/deepseek-r1	DeepSeek-R1

Legacy Claude 3.x names (claude-3-5-sonnet-*, claude-3-opus-*, etc.) are automatically routed to the matching tier via prefix fallbacks.

Override by setting MODEL_CONFIG_PATH=/path/to/your/models.yaml.

---

Environment variables

Variable	Default	Description
NVIDIA_API_KEY	required	nvapi-… key from build.nvidia.com
NVIDIA_BASE_URL	https://integrate.api.nvidia.com/v1	Override for self-hosted NIM
PROXY_HOST	127.0.0.1	Bind address (0.0.0.0 for Docker/remote)
PROXY_PORT	8788	Bind port
PROXY_API_KEY	(unset)	Require clients to send this key as Bearer token
LOG_LEVEL	INFO	DEBUG / INFO / WARNING / ERROR
MODEL_CONFIG_PATH	(bundled)	Path to a custom models.yaml
REQUEST_TIMEOUT_SECONDS	600	Total request timeout (long for reasoning streams)
MAX_RETRIES	2	Upstream retry budget for transient 5xx
RATE_LIMIT_RPM	0 (off)	Per-client requests/minute limit; 0 = disabled
MAX_REQUEST_BODY_MB	0 (off)	Reject bodies larger than this; 0 = unlimited

Variables can also be placed in a .env file in the working directory.

---

API endpoints

Method	Path	Purpose
POST	/v1/messages	Anthropic Messages — streaming & non-streaming
POST	/v1/messages/count_tokens	Approximate token count (cl100k_base)
GET	/v1/models	List model aliases
GET	/healthz	Liveness probe
GET	/metrics	Prometheus metrics (requires [metrics] extra)
POST	/v1/messages/batches	501 stub (not supported by NIM)
POST	/v1/files	501 stub

---

Features

• Streaming — full SSE translation with keepalive ping events
• Tool use — Anthropic tool definitions and results translated to OpenAI function-calling
• Reasoning / thinking — thinking.budget_tokens enforced; <think> tags stripped
• Vision — JPEG/PNG pass-through; GIF/WEBP transcoded to PNG
• PDF documents — base64 PDF blocks extracted to plain text (requires [pdf] extra)
• Model failover — automatic retry on 5xx with next model in the chain
• SIGHUP reload — kill -HUP <pid> reloads models.yaml without restart
• Rate limiting — per-client fixed-window limiter (keyed on metadata.user_id or IP)
• Prometheus metrics — request count, token usage, latency histograms
• Cost estimation — cost_usd_est field in every structured log line

---

Custom model config

Create a models.yaml (start from the bundled default) and set MODEL_CONFIG_PATH:

defaults:
  big: "my-model"
  small: "my-model"

aliases:
  my-model:
    nvidia_id: "org/my-nim-model"
    supports_tools: true
    supports_vision: false
    supports_reasoning: false
    max_context: 131072
    max_output: 16384
    failover_to: []

prefix_fallbacks:
  "claude-": "my-model"

MODEL_CONFIG_PATH=./my_models.yaml nvd-claude-proxy

---

Docker

docker run --rm -p 8788:8788 \
  -e NVIDIA_API_KEY=nvapi-... \
  ghcr.io/khiwn/nvd-claude-proxy:latest

Or clone the repo and use docker compose up.

---

Known limitations

• Prompt caching is silently ignored (NVIDIA NIM has no equivalent).
• thinking.signature is proxy-local — do not forward proxy-generated thinking blocks to the real Anthropic API.
• DeepSeek-R1 + tool use is unreliable; use Nemotron models for agentic workloads.
• Free-tier NIM rate limit is ~40 RPM; heavy tool loops may hit 429s.
• Anthropic server tools (web_search, bash, computer, code_execution, memory) are silently dropped.
• Batch and Files APIs return 501 — NIM has no equivalent.

---

Development

git clone https://github.com/khiwn/nvd-claude-proxy
cd nvd-claude-proxy
make dev    # pip install -e ".[dev,full]"
make test   # pytest
make lint   # ruff + mypy
make run    # uvicorn on :8788

License

MIT — see LICENSE.