OpenPaths

openpaths.io -- AI model gateway -- unified API for chat, image, video, music, speech, embedding, and transcription across 15+ providers.

Requirements: Go 1.25.3, Node.js 22+, npm, PostgreSQL, and gitleaks.

# Clone and enter the repository
git clone https://github.com/lee101/openpaths.git
cd openpaths

# Install JavaScript dependencies and local secrets protection
npm ci
make install-hooks

# Configure local environment
cp .env.example .env
# Edit .env; at minimum set a random JWT_SECRET and DATABASE_URL.

Run the Go service and frontend in separate terminals:

make build
GOMAXPROCS=3 ./bin/openpaths

# Separate terminal
npm run dev

Before opening a pull request, run the same useful checks locally:

make test-go       # Go unit/integration tests
make test-frontend # TypeScript, production build, stable asset checks
make secrets       # Scan the working tree with gitleaks
npm run test:e2e   # Browser tests against the built frontend

The pre-push hook scans the commits in each push. Never commit .env files or real provider credentials. If a credential is detected, revoke/rotate it and remove it from the complete Git history; deleting it from the latest file is not sufficient. SKIP_GITLEAKS=1 git push is reserved for emergencies and does not bypass the GitHub Actions scan.

# Create a local database (adjust the user/password for your machine)
sudo -u postgres psql -c "CREATE USER openpaths WITH PASSWORD 'openpaths';"
sudo -u postgres psql -c "CREATE DATABASE openpaths OWNER openpaths;"

# Configure the application
cp .env.example .env
# Set DATABASE_URL to postgres://openpaths:openpaths@localhost:5432/openpaths?sslmode=disable

# Build and run
make build
GOMAXPROCS=3 ./bin/openpaths

Serves plain HTTP; TLS on https://openpaths.local:9243 is terminated by socat.

# one-time: hosts entry + self-signed cert
grep -q openpaths.local /etc/hosts || echo "127.0.0.1 openpaths.local" | sudo tee -a /etc/hosts
mkdir -p tmp/dev-cert && openssl req -x509 -newkey rsa:2048 \
  -keyout tmp/dev-cert/key.pem -out tmp/dev-cert/cert.pem -days 825 -nodes \
  -subj "/CN=openpaths.local" -addext "subjectAltName=DNS:openpaths.local"

# kill anything on either port and rerun
fuser -k 8092/tcp 9243/tcp; sleep 1; make build &&
  PORT=8092 ./bin/openpaths &>/tmp/openpaths-dev.log &
  socat OPENSSL-LISTEN:9243,fork,reuseaddr,cert=tmp/dev-cert/cert.pem,key=tmp/dev-cert/key.pem,verify=0 TCP4:127.0.0.1:8092 &>/tmp/openpaths-tls.log &

# verify
curl -sk https://openpaths.local:9243/ -o /dev/null

# Requires CUDA 12.x and gobed GPU libs
CUDA_PATH=/usr/local/cuda-12.9 \
GOBED_GPU=/home/lee/code/gobed/gpu \
./scripts/build-gpu.sh

# Run with GPU
./scripts/run-gpu.sh

Or manually:

export LD_LIBRARY_PATH="/home/lee/code/gobed/gpu:/usr/local/cuda-12.9/lib64:$LD_LIBRARY_PATH"
GOMAXPROCS=3 go build -tags="gpu cuda" \
  -ldflags="-extldflags '-Wl,-rpath,/home/lee/code/gobed/gpu -Wl,-rpath,/usr/local/cuda-12.9/lib64'" \
  -o bin/openpaths-gpu ./cmd/openpaths/
./bin/openpaths-gpu

Provider key creation scripts live in rotation/. They create a new supported provider key and update the matching .env value, but intentionally do not revoke the old key.

• POST /v1/chat/completions -- OpenAI-compatible chat
• GET /v1/models -- List available models
• POST /v1/images/generations -- Image generation
• POST /v1/videos/generations -- Video generation
• POST /v1/music/generations -- Music generation
• POST /v1/audio/speech -- Text-to-speech
• POST /v1/audio/transcriptions -- Audio transcription
• POST /v1/embeddings -- Text embeddings
• POST /v1/search -- Search API for Exa and Papers providers
• POST /auth/register -- Register user
• POST /auth/login -- Login
• GET /v1/evals/results -- Live eval snapshot: openpaths/auto vs frontier models across coding, agentic, and creative SVG suites with TTFT, throughput, and cost. Rendered at /evals
• GET /v1/evals/status -- Whether an eval sweep is currently running
• POST /v1/evals/run -- Admin-only. Triggers one eval sweep (cmd/run-evals runs the same sweep from the CLI; cadence via OPENPATHS_EVALS_INTERVAL, disable with OPENPATHS_EVALS_DISABLED=1)
• GET /health -- Health check

POST /v1/chat/completions supports the standard OpenAI-style fields we route internally across providers:

reasoning_effort can be set directly on compatible OpenAI-format requests:

{
  "model": "openpaths/auto-reasoning",
  "messages": [
    {"role": "user", "content": "Design a mesh simplification algorithm."}
  ],
  "reasoning_effort": "high",
  "max_tokens": 2048
}

Use openpaths/auto as the default — OpenPaths picks the backend from your prompt, then orders viable fallback candidates by price unless you pass "routing_strategy": "config". Use openpaths/auto-code, openpaths/auto-fast, openpaths/auto-cheap, openpaths/auto-reasoning, openpaths/auto-vision, or openpaths/auto-image when you want a modality bias. Set reasoning_effort: "auto" on direct thinking models or use openpaths/auto-reasoning to route reasoning depth automatically.

Provider latency, time-to-first-token, and throughput are recorded in usage_logs and surfaced at /stats. Thinking Machines is registered as provider thinkingmachines (Tinker's OpenAI-compatible API), but with no TINKER_API_KEY configured both Inkling ids are served from the open weights on Together: inkling-small (276B/12B active, $0.50/$1.20) and thinkingmachines/inkling (975B/41B, $1.00/$4.05), the latter falling back to or/inkling on OpenRouter. Inference.net is registered as provider inference_net with Nemotron 3 Super, Schematron, ClipTagger, GPT-OSS, Llama, DeepSeek, Qwen, Gemma, and Mistral routes. As of June 29, 2026 its /v1/models response for the configured key did not advertise glm-5.2, so GLM-5.2 remains routed through Z.ai rather than adding a failing Inference.net fallback.

Set reasoning_effort: "auto" on any thinking-capable direct model to keep that model while letting OpenPaths choose none, low, medium, or high from the same embedding table used by auto-think:

{
  "model": "nvidia/deepseek-v4-pro",
  "messages": [
    {"role": "user", "content": "Make a 3D simulation of cogs in a clock."}
  ],
  "reasoning_effort": "auto"
}

The Anthropic-compatible POST /v1/messages endpoint also accepts thinking and output_config.effort, which map onto the same internal reasoning controls. Current Claude models use adaptive thinking and provider-native effort; older Claude models retain budget-based thinking.

Claude Code and Anthropic Agent SDK must use the root URL because they append the Messages path themselves. In PowerShell:

$env:ANTHROPIC_BASE_URL = "https://openpaths.io"
$env:ANTHROPIC_AUTH_TOKEN = $env:OPENPATHS_API_KEY
$env:ANTHROPIC_API_KEY = ""
$env:ANTHROPIC_MODEL = "nvidia/deepseek-v4-pro"
claude

Do not use openpaths.io/v1: it is missing the URL scheme and includes a path those clients append themselves. Anthropic Messages can target OpenAI-compatible models such as nvidia/deepseek-v4-pro; OpenPaths translates messages, tools, tool results, streaming events, and usage. Long stable system prompts and tool definitions get safe automatic prompt-cache breakpoints, while explicit caller cache controls are preserved.

Direct DeepSeek models support thinking: {"type":"enabled"} or {"type":"disabled"} on chat completions. nvidia/deepseek-v4-pro routes to NVIDIA NIM as deepseek-ai/deepseek-v4-pro; OpenPaths sends NVIDIA chat_template_kwargs for thinking mode automatically and can fail over to the direct DeepSeek and Fireworks routes.

npm install
npm run dev

The web UI includes API docs at /docs.