The Problem
AI agent costs are invisible until the bill arrives. Token usage accumulates across parallel sessions and multi-step workflows with no clear signal until something is obviously wrong — or obviously expensive. Existing observability tools require SDK-level lock-in, framework-specific wiring, or a platform subscription before you get any data.
The Solution
TraceStack is a zero-dependency Python SDK that sends traces to a self-hosted FastAPI backend. Install in two minutes, instrument with a decorator or context manager, and get immediate cost and latency data per model — without committing to a platform or changing your agent architecture. Self-hostable for free. Hosted tier starts at $19/mo.
TraceStack is a lightweight LLM execution tracer — a zero-dependency Python SDK that captures every agent call, measures latency, and calculates token costs in real time. Built for developers who need production visibility without adding a heavy observability platform to their stack.
Python SDK published to PyPI. Backend is FastAPI on Render with Turso (libSQL) for the database. Landing page on Cloudflare Pages. Stripe billing with automated key provisioning via AgentMail.
In action
What it does
- Zero-dependency Python SDK — no version conflicts, pure stdlib
- Non-blocking traces — background threads, zero added latency
- Automatic cost estimation for GPT-4o, Claude, Gemini, Llama, and more
- Decorator, context manager, and manual span APIs
- REST API with stats, trace history, and per-model breakdowns
- Self-hostable FastAPI backend — deploys to Render in minutes
- Turso-backed — persistent, managed SQLite at the edge
See it live.
Open TraceStack →More work
Archalist
SaaS · Design
AI-curated design directory focused on enduring tools, methods, and references.
View →KerfOS
SaaS · Woodworking
Cabinet design software for the CNC woodworker.
View →Forge
SaaS · Hardware
AI-assisted design for physical things — from concept to cut file.
View →