provekit 0.6.0

Drop-in agent tracing — add one decorator, get a project key, review every run your agent makes.

ProveKit

One decorator. The whole agent flow. Add one @pk.trace at your agent's entrypoint and review every run — the model calls, tools, and steps — as a nested flow in your ProveKit portal. OpenAI/Anthropic calls capture themselves. Open source, self-hostable, no framework lock-in.

Install

pip install "provekit[trace]"

Use

import provekit.trace as pk

# .env
#   PROVEKIT_API_KEY=pk_...        (create a project + key in the portal)
#   PROVEKIT_ENDPOINT=https://your-provekit-host

@pk.trace(name="support-agent")
def run_agent(question: str) -> str:
    docs = retrieve(question)          # wrap sub-steps with `with pk.span("retrieve"):`
    return chat(question, docs)        # the OpenAI/Anthropic call captures itself

Run your agent, open Traces in the portal, and every run shows up as a nested waterfall with per-span input, output, and token usage.

It's OpenTelemetry under the hood, so your data is portable and nothing is locked in — but you never have to touch OTel. Fail-open by design: if the key/endpoint are unset or the portal is unreachable, your app runs completely unaffected.

Run the portal (self-host)

The web app + ingest server ship in the same package:

pip install "provekit[server]"
uvicorn provekit.main:app --port 8100

See the repo for Docker/compose and the full tracing guide.

• Repository: https://github.com/MobirizerServices/ProveKit
• License: MIT