Wont

Wont watches how your AI agents actually work, finds the steps they repeat, and shows which of that work is reusable. A wont is a habitual way of doing something; that is what Wont looks for in agent behavior.

Status: early release (0.1.0). Today Wont observes and measures. It does not execute anything in place of your agent. Every number it shows is a measured fact about your own traffic; it makes no savings estimates.

What it does

• Observe. A small Python SDK records each agent run as a structured trace: which tools were called, in what dependency order, with what side-effect class. It stores the shape of values and keyed hashes, not the values themselves.
• Find repetition. Wont reconstructs the data flow between steps and groups runs that perform the same procedure.
• Show the lifecycle. Each repeated procedure is observed, a candidate, qualified, or deoptimized, derived from evidence you can inspect. Nothing is "active": Wont does not run procedures for you.
• Measure. Cost and latency appear only when your traces report them, and are marked unavailable otherwise.

Install

pip install wont

The package is small: it depends only on pydantic. The server and dashboard ship in the source repository and the container image; pip install "wont[server]" adds the libraries they need.

Use it

import os

from wont import Wont

wont = Wont(
    api_key=os.environ["WONT_API_KEY"],      # a project key from the dashboard (wnk_...)
    base_url=os.environ["WONT_BASE_URL"],
)

with wont.execution(workflow_version="refund-status") as run:
    result = support_agent.run(request)       # your agent, unchanged

For tools you call yourself, observe_tool records a call without ever raising into your agent. Instrumentation failures are swallowed by design; an observation problem must never break a run. See the SDK guide.

Coding agents

If your agent is Claude Code, wont claude-code install prints a hook block that records sessions locally first, with an explicit step before anything is sent. See the recorder guide and the pilot runbook.

Safety stance

Wont is conservative by construction: when it is unsure whether two steps are independent, it keeps them ordered; a false fallback is acceptable, silent incorrect execution is not. Project API keys are agent credentials and are never used in a browser. The dashboard signs people in with their own sessions. See data handling.

Run the server

From a source checkout (Python 3.12 or newer, Docker for PostgreSQL, Node for the dashboard build):

pip install -e ".[server]"
(cd apps/web && npm ci && VITE_WONT_DATA_SOURCE=api npm run build)   # once; builds the dashboard
docker compose up -d postgres
wont serve --dev

wont serve --dev is for local use only: it listens on 127.0.0.1, uses the local development database, applies pending migrations, serves the dashboard at http://127.0.0.1:8000, and prints sign-in links in the console (there is no email in development). It refuses to run with WONT_ENV=production. For a real deployment use the container image and the reference deployment in deploy/; see the operations runbook.

Documentation

• Getting started: sign in to your first data
• API keys: lifecycle, rotation, and what a key can do
• Data controls: what is stored, retention, export, deletion
• Quickstart
• Python SDK
• Claude Code recorder
• Naming and conventions

License

Apache License 2.0. See LICENSE.

To report a vulnerability, see SECURITY.md.