How I shipped an open-source AI tool in a week

Over seven days I built and published Archon, an open-source control plane for AI agents: a stdlib-only Python service, a single-file dashboard, and a constrained control API. This is the build log.

The architecture: observe and control

The system exposes two surfaces with different trust requirements:

Observe — an append-only event stream, one JSON contract for every framework.
Control — a constrained API that acts on running agents.

Data flows one way, control the other:

logs → adapter → event stream → dashboard      (observe)
dashboard → control API → executor             (act)

An adapter normalises a framework's native logs into the shared event schema. Adding a framework is one adapter file, not a rewrite.

The safety model: capability-based, not command-based

A control plane that can start and stop agents is a dangerous primitive. So the control surface has no code-execution path.

Actions are an allowlist, defined statically in a host config file. The API accepts a single shape:

{ "action": "restart", "target": "trading-monitor" }

and resolves it against that table. There is no eval, no subprocess with client-derived argv, no shell interpolation. A client can only select from a pre-reviewed set of actions — never construct one. The protocol's grammar is the sandbox.

The stack: zero dependencies

The server is stdlib (http.server / socketserver); the UI is one HTML file; there are no third-party imports. That's deliberate: a control plane is incident-path software — it has to boot on a cold host, with no network and no build step, when something is already on fire.

The packaging failure that nearly shipped a broken wheel

Build green, tests green, wheel uploaded cleanly — and the dashboard file wasn't in it. setuptools only packages modules by default; a non-.py asset must be declared explicitly:

[tool.setuptools.package-data]
archon = ["dashboard.html"]

python -m zipfile -l dist/*.whl is now a mandatory release step. Inspect the artifact, not just the build.

Distribution naming

archon was already registered on PyPI, so I published under archon-hq — distribution name and import name are decoupled in Python packaging: the import stays archon, the entry point stays archon serve.

pip install archon-hq
archon serve --sessions-dir ~/.openclaw-autoclaw/agents/main/sessions \
             --control-config examples/control-config.json

The demo GIF

I drove Playwright against system Chrome (channel="chrome" — the bundled Chromium was unavailable, and raw headless Chrome collided with a WebView2 runtime) and assembled the frames into a GIF with Pillow. A 30-second watchable demo converts better than any paragraph.

What I'd change

Inspect the artifact. The wheel bug cost an hour that one zipfile -l would have saved.
Run name clearance first. A trademark check in the relevant classes surfaced near-identical marks in the same vertical — cheaper to learn on day six than month six.
Write the README as a spec, from day one. Retrofitting it meant rediscovering my own contract.

What's next

Archon is pre-alpha but running in production against my own agent fleet. Next: real users, a hosted tier, a trading edition.

A week is enough to ship something real — if the constraint is real.