Keeto

Zero-config AI observability for Python. Automatic tracing, cost tracking, and dashboards for every LLM call — plus a @trace decorator to instrument your own pipelines.

from keeto import monitor
monitor.start()

# Your existing code — completely unchanged
response = client.chat.completions.create(model="gpt-4o", messages=[...])

monitor.dashboard()

┌─────────────────────────────────────────────────────────────┐
│  Traces: 42   Cost: $0.0312   Tokens: 18,400   Errors: 0   │
├───────────────┬──────────┬────────┬─────────┬──────────────┤
│ Model         │ Calls    │ P50    │ P95     │ Cost         │
├───────────────┼──────────┼────────┼─────────┼──────────────┤
│ gpt-4o        │ 28       │ 843ms  │ 2.1s    │ $0.0289      │
│ gpt-4o-mini   │ 14       │ 312ms  │ 680ms   │ $0.0023      │
└───────────────┴──────────┴────────┴─────────┴──────────────┘

Install

pip install keeto

# SQLite persistence + interactive TUI
pip install "keeto[sqlite,tui]"

# Browser dashboard
pip install "keeto[sqlite,web]"

Two ways to use Keeto

1. Auto-detect AI SDK calls

Drop two lines into your app. Keeto scans installed packages and patches their HTTP transport — no wrappers, no code changes.

from keeto import monitor
monitor.start()

# OpenAI, Anthropic, LangChain, Gemini … all captured automatically

keeto: loaded plugins → openai, anthropic

2. Instrument your own pipeline

Use @trace to add stage-by-stage timing to RAG pipelines, multi-step workflows, or any custom framework:

from keeto import monitor, trace

monitor.start()

@trace("embedding")
def embed(text: str) -> list[float]: ...

@trace("retrieval")
def search(vec: list[float]) -> list[str]: ...

@trace("rerank")
def rerank(docs: list[str], query: str) -> list[str]: ...

@trace("llm")
def generate(docs: list[str]) -> str: ...

# Run your pipeline
answer = generate(rerank(search(embed("What is the capital of France?")), "..."))

monitor.pipeline_breakdown()

Trace a3f8bc12

embedding        18 ms
retrieval        12 ms
rerank           65 ms
llm            1100 ms
──────────────────────
Total          1195 ms

Slowest stage: llm (92.1%)

@trace works on sync and async functions. Common stage names (embedding, retrieval, llm, rerank, search) are mapped to the correct span kind automatically.

Features

Integrations

Keeto auto-detects whichever SDKs you have installed:

• OpenAI — chat, embeddings, tools, streaming
• Anthropic — messages, tools, streaming
• LangChain — chains, agents, retrievers
• LlamaIndex — query engines, agents
• LiteLLM — all providers via callback
• Google Gemini — generate_content, streaming
• Ollama — local models
• OpenAI Agents SDK — traces, tool calls
• PydanticAI — agents, tools
• FastAPI — per-request trace context middleware
• vLLM — local inference server
• Custom / own framework — @trace decorator

Configuration

from keeto import Monitor
from keeto.storage.sqlite import SQLiteStorage

monitor = Monitor(
    storage=SQLiteStorage("./keeto.db"),   # persist across restarts
    sample_rate=0.1,                        # capture 10% in production
    scrub_pii=True,                         # redact PII before storage
    store_prompts=True,                     # set False for metadata-only
)
monitor.start()

monitor.set_budget(daily_usd=10.0, session_usd=2.0)
monitor.set_token_budget(monthly=1_000_000)

Manual span context

For fine-grained control, use monitor.span() directly to group stages under a named root:

with monitor.span("rag-pipeline") as ctx:
    ctx.set_attribute("query", query)
    vec   = embed(query)
    docs  = search(vec)
    answer = generate(docs, query)

Dashboard modes

monitor.dashboard()           # Rich table in terminal (no extra deps)
monitor.dashboard("tui")      # Interactive TUI — requires keeto[tui]
monitor.dashboard("web")      # Browser dashboard — requires keeto[web]

CLI

keeto traces                  # list recent traces
keeto dashboard               # launch TUI
keeto analyze                 # print recommendations
keeto export --format json    # export to file
keeto replay <trace-id>       # re-send a captured request
keeto doctor                  # diagnose setup issues

Links

• Documentation
• Quickstart
• Custom Instrumentation (@trace)
• PyPI
• Changelog

License

MIT