OpenTracy

Drop-in OpenAI-compatible SDK. Every request becomes a trace; traces become datasets; datasets become distilled custom models; the routing layer swaps those models in under your app via aliases — so your cost curve goes down over time without code changes.

Sponsors

Try it in Colab (no install)

Each notebook runs end-to-end on a free Colab runtime — bring your own OpenAI key, optionally Anthropic / Groq.

#	Notebook	One-line pitch	Colab
01	Quickstart	First completion() call, see _cost + _latency_ms, swap providers
02	Drop in over the OpenAI SDK	Keep from openai import OpenAI, change only base_url
03	Semantic auto-routing	One prompt, the right model of 13 — learned, not rule-based
04	Ticket classifier (real app)	End-to-end support-ticket classifier with cost breakdown
05	Distillation — train your student	Turn trace history into a distilled tiny model
06	Serve your distilled model	Four serving paths from load-the-adapter to alias swap

Colab heads-up — traces only show up in the dashboard if you set OPENTRACY_ENGINE_URL before import opentracy. Every notebook has a commented-out cell at the top with the two lines you need.

Install

pip install opentracy

Quick start

import opentracy as lr

resp = lr.completion(
    model="openai/gpt-4o-mini",
    messages=[{"role": "user", "content": "Hello"}],
)
print(resp.choices[0].message.content)
print(f"cost: ${resp._cost:.6f}  latency: {resp._latency_ms:.0f}ms")

Works with 13 providers out of the box: OpenAI, Anthropic, Gemini, Groq, Mistral, DeepSeek, Together, Fireworks, Cerebras, Sambanova, Perplexity, Cohere, Bedrock.

Connecting to the OpenTracy platform (traces, dashboards, distillation)

By default lr.completion() goes direct to the provider, so calls do not appear in the OpenTracy dashboard. To route every call through a running engine — the only way traces, metrics, and the distillation loop get data — set OPENTRACY_ENGINE_URL before importing the SDK:

import os
os.environ["OPENTRACY_ENGINE_URL"] = "http://<your-opentracy-host>:8080"  # engine port
import opentracy as lr

resp = lr.completion(
    model="openai/gpt-4o-mini",
    messages=[{"role": "user", "content": "Hello"}],
)
# trace now visible in the dashboard on the UI host at :3000

Alternatives:

• Per-call: pass force_engine=True, api_base="http://<host>:8080/v1" to lr.completion(...).
• Drop-in OpenAI SDK (no code change beyond base_url — see below).

API keys for your providers should be saved once via the UI (Settings → API Keys) or the API (POST /v1/secrets/<provider>); the engine picks them up immediately from ~/.opentracy/secrets.json.

Routing with fallbacks

router = lr.Router(
    model_list=[
        {"model_name": "smart", "model": "openai/gpt-4o"},
        {"model_name": "smart", "model": "anthropic/claude-sonnet-4-6"},
    ],
    fallbacks=[{"smart": ["deepseek/deepseek-chat"]}],
)
resp = router.completion(model="smart", messages=[{"role": "user", "content": "Hi"}])

Drop-in replacement for the OpenAI SDK

Point any existing OpenAI app at the OpenTracy engine — zero code changes beyond base_url:

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8080/v1", api_key="any")
# All 13 providers routed through the OpenTracy engine; every request is a trace.

Distillation — what makes OpenTracy different from a plain gateway

from opentracy import Distiller

d = Distiller()
# Submit a dataset built from your own traces, pick a teacher + student model,
# and OpenTracy trains the distilled model and serves it behind a routing alias
# you can point traffic at. Your app code never changes.

Install the training extras for the distillation pipeline:

pip install opentracy[distill]

Self-host the full platform (traces + UI + REST API)

git clone https://github.com/OpenTracy/opentracy.git
cd opentracy
make up   # Gateway + ClickHouse analytics + Python API + UI

Engine at http://localhost:8080, Python API at http://localhost:8000, UI at http://localhost:3000.

GPU is optional

The Python API container runs on CPU by default — no nvidia-container-toolkit required, so it works on plain cloud VMs and laptops without a GPU. Local training and inference paths fall back to CPU automatically.

To enable GPU acceleration (faster distillation training, local inference):

DOCKER_RUNTIME=nvidia docker compose up -d

This requires the NVIDIA Container Toolkit on the host. Persist the variable in .env next to docker-compose.yml if you always run with GPU.

What OpenTracy Does

Requests ──► Gateway (13 providers) ──► Traces (ClickHouse)
                                            │
                                    ┌───────┴───────┐
                                    ▼               ▼
                              Clustering        Analytics
                              (domains)        (cost/latency)
                                    │
                              ┌─────┴─────┐
                              ▼           ▼
                         Evaluations   Distillation
                        (AI metrics)  (training data)

1. Route — proxy to 13 LLM providers with fallbacks, retries, and cost tracking
2. Observe — every request/response stored in ClickHouse with full content
3. Cluster — auto-group prompts by domain using embeddings + LLM labeling
4. Evaluate — run models against domain datasets with built-in and AI-suggested metrics
5. Distill — export input/output pairs per domain for fine-tuning smaller models

Features

Gateway

• 13 LLM Providers through one OpenAI-compatible API
• Python SDK — lr.completion() one-liner
• Router Class — load balancing, fallbacks, retries, 4 strategies
• Streaming — all providers including Anthropic & Bedrock SSE translation
• Cost Tracking — 70+ models with per-token pricing on every response
• Vision / Multimodal — images via base64 or URL
• Tool Calling — function calls with cross-provider translation
• Semantic Routing — auto-select the best model per prompt (with weights)

Observability

• ClickHouse Analytics — traces, cost, latency, model-level stats
• Full Content Capture — input/output text stored for every request
• Trace Scanning — AI agent detects hallucinations, refusals, quality regressions
• Real-time Dashboard — UI with filters, search, trace detail drawer

Domain Clustering

• Auto-clustering — groups prompts by semantic similarity (KMeans + MiniLM embeddings)
• LLM Labeling — AI agent names each cluster (e.g., "JavaScript Concepts", "Business Strategy")
• Quality Gates — coherence scoring, outlier detection, merge suggestions
• Input + Output Storage — full pairs stored per cluster for distillation

Evaluations

• Run Evaluations — send dataset samples through models, score and compare
• 6 Built-in Metrics — exact match, contains, similarity, LLM-as-judge, latency, cost
• AI Metric Suggestion — harness agent analyzes dataset domain and creates tailored metrics
• Background Execution — evaluations run async with progress tracking
• Model Comparison — side-by-side results with winner determination

Distillation

• BOND Pipeline — teacher → LLM-as-Judge curation → LoRA training (Unsloth) → GGUF export
• Dataset Support — use domain clusters or custom datasets as training source
• UI + API — create and monitor jobs via dashboard or REST endpoints

Harness (AI Agent System)

• Agent Runner — loads .md agent configs, calls LLM, parses structured output
• 7 Agents — cluster labeler, coherence scorer, outlier detector, merge checker, trace scanner, eval generator, metrics suggester
• Memory Layer — persistent agent memory with query/summary
• Tool Access — agents can call tools (list traces, query datasets, etc.)

Supported Providers

Installation

pip install -e ".[openai,anthropic,api]"   # SDK + common providers
pip install -e ".[all]"                     # everything
pip install -e ".[train]"                   # training/distillation deps (CUDA)

Python SDK

Completion

import opentracy as lr

response = lr.completion(
    model="openai/gpt-4o-mini",
    messages=[{"role": "user", "content": "Hello!"}],
)

# Streaming
for chunk in lr.completion(model="openai/gpt-4o-mini", messages=[...], stream=True):
    print(chunk.choices[0].delta.content or "", end="")

# Fallbacks
response = lr.completion(
    model="openai/gpt-4o-mini",
    messages=[...],
    fallbacks=["anthropic/claude-haiku-4-5-20251001", "groq/llama-3.3-70b-versatile"],
    num_retries=2,
)

Router (Load Balancing)

router = lr.Router(
    model_list=[
        {"model_name": "smart", "model": "openai/gpt-4o"},
        {"model_name": "smart", "model": "anthropic/claude-sonnet-4-20250514"},
        {"model_name": "fast",  "model": "groq/llama-3.3-70b-versatile"},
    ],
    fallbacks=[{"smart": ["deepseek/deepseek-chat"]}],
    strategy="round-robin",  # or: least-cost, lowest-latency, weighted-random
)

response = router.completion(model="smart", messages=[...])

Drop-in OpenAI SDK

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8080/v1", api_key="any")
client.chat.completions.create(model="openai/gpt-4o-mini", messages=[...])
client.chat.completions.create(model="anthropic/claude-haiku-4-5-20251001", messages=[...])
client.chat.completions.create(model="mistral/mistral-small-latest", messages=[...])

Running

API Keys

Configure via the UI, environment variables, or ~/.opentracy/secrets.json:

export OPENAI_API_KEY=sk-...
export ANTHROPIC_API_KEY=sk-ant-...
make start-full

API Endpoints

Gateway (Go Engine — port 8080)

Python API (port 8000)

Architecture

go/                              # Go engine (high-performance gateway)
├── cmd/opentracy-engine/            # Entry point
├── internal/
│   ├── provider/                # 13 providers
│   ├── server/                  # HTTP handlers + session management
│   ├── clickhouse/              # Trace writer + 8 migrations
│   ├── router/                  # UniRoute algorithm + LRU cache
│   └── embeddings/              # ONNX MiniLM embedder

opentracy/                    # Python layer (analytics, clustering, evals)
├── api/server.py                # FastAPI — analytics, clustering, evaluations, metrics
├── sdk.py                       # completion(), acompletion(), Router class
├── clustering/
│   ├── pipeline.py              # Extract → embed → cluster → label → store
│   ├── labeler.py               # LLM-powered cluster labeling via harness
│   └── quality.py               # Coherence, diversity, noise quality gates
├── harness/
│   ├── runner.py                # Agent executor (JSON parsing, retry, tools)
│   ├── tools.py                 # Agent tools (query traces, datasets, etc.)
│   ├── memory_store.py          # Persistent agent memory
│   └── agents/                  # 7 agent configs (.md files)
│       ├── cluster_labeler.md
│       ├── coherence_scorer.md
│       ├── outlier_detector.md
│       ├── merge_checker.md
│       ├── trace_scanner.md
│       ├── eval_generator.md
│       └── metrics_suggester.md
├── distillation/
│   ├── pipeline.py              # 4-phase orchestrator (data gen → curation → train → export)
│   ├── data_gen.py              # Teacher model candidate generation
│   ├── curation.py              # LLM-as-Judge scoring & selection
│   ├── trainer.py               # SFT/BOND fine-tuning (Unsloth + LoRA)
│   ├── export.py                # LoRA merge + GGUF conversion
│   ├── repository.py            # ClickHouse persistence
│   ├── router.py                # API endpoints
│   └── schemas.py               # Pydantic models & model catalog
├── evaluations/                 # Evaluation runs & results
├── datasets/                    # Dataset CRUD, from-traces, auto-collect
├── metrics/                     # Metric definitions & validation
├── experiments/                 # A/B experiments & comparison
├── annotations/                 # Human annotation queues
├── auto_eval/                   # Automated evaluation configs & triggers
├── eval_agent/                  # AI-powered eval setup assistant
├── proposals/                   # Decision engine proposals
├── trace_issues/                # Issue scanning & detection
├── training/                    # Custom router training (UniRoute)
├── storage/
│   ├── clickhouse_client.py     # Analytics queries
│   ├── secrets.py               # API key management
│   └── state_manager.py         # File-based state persistence
├── model_prices.py              # 70+ models with pricing
└── mcp/                         # Claude Code MCP server

ui/                              # React dashboard
├── src/features/
│   ├── traces/                  # Trace explorer with drawer, filters, timeline
│   ├── evaluations/             # Run evaluations, metrics, experiments
│   └── distill-dataset/         # Dataset management, clustering, export

Evaluation Workflow

# 1. Send traffic through the gateway (traces auto-captured)
curl http://localhost:8080/v1/chat/completions \
  -d '{"model": "openai/gpt-4o-mini", "messages": [{"role": "user", "content": "Explain closures in JS"}]}'

# 2. Run clustering to group prompts by domain
curl -X POST http://localhost:8000/v1/clustering/run?days=30&min_traces=5

# 3. AI suggests metrics for a domain dataset
curl -X POST http://localhost:8000/v1/auto-eval/suggest-metrics \
  -d '{"dataset_id": "cluster:run-id:2"}'

# 4. Run evaluation comparing models on that dataset
curl -X POST http://localhost:8000/v1/evaluations \
  -d '{"name": "JS eval", "dataset_id": "cluster:run-id:2",
       "models": ["openai/gpt-4o-mini", "mistral/mistral-small-latest"],
       "metrics": ["similarity", "latency", "cost", "llm_judge"]}'

# 5. Check results
curl http://localhost:8000/v1/evaluations/{id}/results

Distillation

BOND-style distillation pipeline: generate candidates with a teacher model, score them with LLM-as-Judge, fine-tune a student model with LoRA, and export to GGUF.

make install-train   # install training deps (requires CUDA)

Via UI at http://localhost:3000 → Distillation, or via API:

curl -X POST http://localhost:8000/v1/distillation/default/jobs \
  -H "Content-Type: application/json" \
  -d '{
    "teacher_model": "openai/gpt-4o-mini",
    "student_model": "unsloth/Llama-3.2-1B-Instruct-bnb-4bit",
    "num_candidates": 5,
    "dataset_id": "my-dataset"
  }'

Semantic Routing

With pre-trained weights, the router picks the best model per prompt:

make download-weights   # download from HuggingFace
make gateway-router     # start with semantic routing enabled

router = load_router()
decision = router.route("Explain quantum computing")
print(f"Best model: {decision.selected_model}")
print(f"Expected error: {decision.expected_error:.4f}")

Training Custom Routers

from opentracy import full_training_pipeline, TrainingConfig, PromptDataset, create_client

train_data = PromptDataset.load("train.json")
val_data = PromptDataset.load("val.json")

clients = [
    create_client("openai", "gpt-4o"),
    create_client("openai", "gpt-4o-mini"),
    create_client("groq", "llama-3.1-8b-instant"),
]

result = full_training_pipeline(
    train_data, val_data, clients,
    TrainingConfig(num_clusters=100, output_dir="./weights"),
)

MCP Integration (Claude Code)

pip install opentracy[mcp]

Add to ~/.claude/settings.json:

{
  "mcpServers": {
    "opentracy": {
      "command": "python",
      "args": ["-m", "opentracy.mcp"]
    }
  }
}

Tools: opentracy_route, opentracy_generate, opentracy_smart_generate, opentracy_list_models, opentracy_compare.

Development

make help               # show all commands
make install            # install Python SDK + Go deps
make install-train      # install training/distillation deps (CUDA)
make up                 # start full local stack (ClickHouse + engine + API + UI)
make stop               # stop all local services
make test               # run all tests
make lint               # lint all code

License

MIT License - see LICENSE file for details.

Links

• GitHub Repository
• HuggingFace Weights