Claude-NIM Proxy
Use 100+ NVIDIA NIM models with Claude Code
Translates Anthropic API to OpenAI-compatible format — zero config changes.
Model Family Selection
Claude Code Running Through the Claude-NIM Proxy
Claude Code Launched via Proxy with Session Stats
Claude Code Native /model Picker with FCC Gateway Models
Requirements
- VS Code 1.80+
- NVIDIA NIM API key — free at build.nvidia.com
- Claude Code CLI — auto-installed on first use
Quick Start
VS Code Extension
- Install from the Marketplace or Open VSX
- Run
Claude NIM Proxy: Manage NVIDIA NIM API Key to set your key
- Run
Claude NIM Proxy: Launch Claude Code with Proxy or press Ctrl+Shift+Alt+N
CLI (npm)
# Install globally
npm install -g claude-nim
# Run
claude-nim # Interactive terminal UI
claude-nim --model deepseek-ai/deepseek-r1 # Explicit model
claude-nim --port 8080 --api-key nvapi-xxx # Custom port + key
claude-nim --serve-only --port 3456 # Proxy server only (no Claude Code)
claude-nim --version # Show version
claude-nim --help # All options
CLI (bunx — no install)
bunx --yes claude-nim # Interactive setup
bunx claude-nim --model deepseek-r1 # Explicit config
How It Works
Claude Code ──→ Claude-NIM Proxy ──→ NVIDIA NIM API
(Anthropic API) (localhost:3456) (OpenAI-compatible)
- Install the extension and set your API key
- Start the proxy from the status bar or command palette
- Launch Claude Code through the proxy
- Stop the proxy — everything reverts, zero permanent changes
Why Claude-NIM Proxy?
|
Claude-NIM Proxy |
CLI Proxies |
CCProxy |
Claude Code Router |
| VS Code integration |
Status bar, commands, SecretStorage, model browser |
None |
CLI only |
None |
| Security |
Prompt injection scrubbing, context pruning, AES-256-GCM keys |
None |
None |
None |
| Setup |
One command or extension install |
Manual env vars + config files |
Binary + config |
npm install + ccr start |
| Model routing |
Gateway IDs, 100+ NIM catalog, /model picker |
Generic passthrough |
Generic |
Generic |
| Language |
TypeScript, zero runtime deps |
Python/Go |
Go |
Node.js + YAML |
| Tests |
104+ unit tests + stress tests |
Minimal |
None public |
Minimal |
| Live settings |
Port/timeout/cache apply without restart |
Requires restart |
Requires restart |
Requires restart |
Features
Model Router & Gateway
- Gateway model IDs — FCC-compliant
anthropic/nvidia_nim/<modelId> format
- Native
/model picker — All NIM models appear in Claude Code's model selector
- Real-time switching — Change models without restarting
- 100+ models — DeepSeek, Llama, Qwen, Mistral, Gemma, Phi, Nemotron, and more
Full Anthropic Content Translation
| Content Type |
Handling |
text / tool_use / tool_result |
Full conversion to OpenAI format |
| Image (base64) |
Converted to image_url data URI |
| Mixed text + tools |
Split into separate messages |
system prompt (string/array) |
Converted to system message |
tool_choice (auto/any/tool) |
Full mapping |
Security
- Prompt injection scrubbing — Neutralizes
ignore previous instructions patterns
- Context pruning — Auto-trims large tool outputs (>100K chars)
- 10 MB body limit — Prevents memory exhaustion
- Unicode sanitization — Strips encoding corruption characters
- Localhost-only binding — Never exposed to network
VS Code Commands
| Command |
Shortcut |
Description |
| Toggle Proxy Server |
Ctrl+Shift+N |
Start or stop the proxy |
| Launch Claude Code |
Ctrl+Shift+Alt+N |
Open pre-configured terminal |
| Manage API Key |
— |
Set, update, or clear your key |
| Select Default Model |
— |
Browse 100+ NIM models |
| Toggle Debug Logging |
— |
Enable/disable debug output |
| Toggle Show Reasoning |
— |
Show/hide <think> output |
Configuration
| Setting |
Default |
Description |
nvidia-nim.proxyPort |
3456 |
Proxy server port |
nvidia-nim.defaultModel |
"" |
Default model (empty = require in request) |
nvidia-nim.modelsCacheTTL |
5 |
Model cache TTL in minutes |
nvidia-nim.requestTimeout |
120 |
Stream idle timeout in seconds |
Dashboard
Open http://127.0.0.1:3456/dashboard for:
- Real-time request visualization
- Live stats (tokens, latency, throughput)
- Model browser and switcher
- Request history with filtering
- Live log stream
API Endpoints
| Endpoint |
Method |
Description |
/api/model |
GET/POST |
Get or set current model |
/api/models |
GET |
List available NIM models |
/api/key |
POST |
Update API key |
/api/metrics |
GET |
SSE stream of real-time metrics |
/api/stats |
GET |
Aggregate stats |
Standalone CLI
claude-nim # Interactive terminal UI
claude-nim --model deepseek-ai/deepseek-r1 # Explicit model
claude-nim --port 8080 --api-key nvapi-xxx # Custom port + key
claude-nim --serve-only --port 3456 # Proxy only (no Claude Code launch)
Features: encrypted key storage, dynamic port selection, Claude auto-detection, zombie-free teardown. Requires Bun runtime.
Error Handling
- HTTP status mapping (401, 429, 503)
- SSE error events in stream
- VS Code error notifications
- Exponential backoff retry with
Retry-After support
- Configurable stream idle timeout
- 10 MB body size limit
- JSON parse error messages with context
Supported Models
Any model on build.nvidia.com:
DeepSeek R1/V3/V4, Llama 3.x/4.x, Mistral Large/Medium, Qwen 3/2.5, Kimi K2.x, Nemotron Ultra/Super, Gemma 3, Phi 4, Command-R+, and 100+ more.
Build
bun install
bun run compile # Build → out/
bun run test # 104+ unit tests
bun run lint # ESLint
bun run package:vsix # Package VS Code extension
bun run build:exe:win # Windows standalone binary
bun run build:exe:linux # Linux standalone binary
bun run build:exe:mac # macOS standalone binary
Contributing
Issues and PRs welcome at github.com/claude-server/claude-nim.
License
MIT — see LICENSE.
Author
Rithika Liyanage — github.com/k-rithik04
