dottle-sdk 0.1.2

Dottle SDK — Instrument your AI agents in 3 lines of code. See every LLM call, tool, cost, and failure in real time.

Dottle SDK

Monitor your AI agents in minutes. See every LLM call, tool use, cost, latency, and failure — in real time.

dottle.dev · Dashboard · Docs

---

Install

pip install dottle-sdk

Quickstart

import dottle

# 1. Configure once at startup
dottle.configure(api_key="dtl_live_...")

# 2. Wrap your agent run in a session
with dottle.session("my-agent", user_id="user_123") as sid:

    # 3. Track each LLM call as a span
    with dottle.span("llm", "gpt-4o reply") as s:
        response = openai_client.chat.completions.create(...)
        s.record_tokens(
            prompt_tokens=response.usage.prompt_tokens,
            completion_tokens=response.usage.completion_tokens,
            model="gpt-4o",
        )

That's it. Open app.dottle.dev to see your agent's sessions, costs, and errors live.

---

Get your API key

1. Sign up at app.dottle.dev
2. Create an organization → create a project
3. Copy the dtl_live_... key from Project Settings

---

What gets tracked

Signal	How
LLM calls	dottle.span("llm", ...) + s.record_tokens(...)
Tool calls	dottle.span("tool", ...)
Errors	s.record_error(exc) or automatic on exception
Cost	Calculated from token counts + model
Latency	Automatic (start/end of each span)
User	Pass user_id / user_email to dottle.session()

---

Full example with Anthropic

import anthropic
import dottle

dottle.configure(api_key="dtl_live_...")
client = anthropic.Anthropic()

def run_agent(user_message: str, user_email: str):
    with dottle.session("support-agent", user_email=user_email) as sid:
        with dottle.span("llm", "claude-3-5-sonnet") as s:
            response = client.messages.create(
                model="claude-3-5-sonnet-20241022",
                max_tokens=1024,
                messages=[{"role": "user", "content": user_message}],
            )
            s.record_tokens(
                prompt_tokens=response.usage.input_tokens,
                completion_tokens=response.usage.output_tokens,
                model="claude-3-5-sonnet-20241022",
            )
        return response.content[0].text

---

Zero performance impact

All calls are fire-and-forget (background thread). Your agent never waits for Dottle. If Dottle is unreachable, your agent keeps running — monitoring failures are silently swallowed.

---

License

MIT