Codel

What was appealing here was the interface: a terminal, a browser and a file editor in one self-hosted web page, so you could watch an agent work without a subscription or an account. That idea was ahead of its time in early 2024 and it is table stakes now.

Nothing has landed since, and an agent that does not track model behaviour goes stale faster than almost any other kind of software. Pick OpenHands, which does the same job with people still working on it, and read this one if you want a compact example of the architecture.

The failure mode is drift. Base images accumulate vulnerabilities, pinned Python and Node dependencies age out of support, and provider APIs deprecate the shapes this code was written against. None of that produces a clean error; it produces a run that fails in a way nobody will explain to you, because there is nobody.

Anyone running it should treat the deployment as a fixed artefact and pin everything. What it does right: each task gets its own container, so a destructive command is bounded by a lifetime measured in one task rather than reaching the host.

Most agents of this generation logged to a scrolling pane and forgot everything on exit. Persisting every command and its output to a relational store means a run can be queried after the fact, compared with another, and used as evidence about what the model actually did rather than what a transcript suggested. That is the beginning of a reproducibility story.

It stops there. Nothing in the design defines a success criterion, so the record supports post-hoc analysis and not automated verification. No benchmark is reported, and the documentation is a README.

There is no company, no funding and nothing to acquire, which makes this useful mainly as a data point about the category. An independent harness built on somebody else's model has no defensible position once that model's vendor ships an equivalent, and the vendors all did. The engineering was fine and the strategic ground disappeared underneath it.

That pattern repeats across this board and it is the single most reliable predictor of which projects go quiet. Position: none. The lesson is worth more than the code, and the code is free anyway.

Standing this up means my platform team operates a web application that creates containers on demand and stores an API key, from a codebase nobody patches. There is no identity integration, so access control would be whatever network boundary we build around it, and no audit surface beyond what we bolt on ourselves.

That is a permanent maintenance obligation acquired in exchange for a capability we can buy supported. There is no vendor, no agreement and no security contact. Not yet, and I would decline it again next quarter.

Starting it is a single docker run mapping 3000 to the container, with the key in an environment variable, and pointing it at a local Ollama server instead means it works with no account anywhere. That is the configuration I want from every tool and get from very few.

AGPL-3.0 keeps a fork honest, and a fork is the only path this has left. The code is small enough that adopting it is realistic for one person, which is the compliment I pay to abandoned software I still respect. I have kept a copy. I have not started the fork.