Skip to content

Aldwin 0.0.4 - any model, including local

Why Aldwin dropped OpenAI lock-in, how local-first providers work, and why this release matters.

Aldwin started life with one cloud path: OpenAI. That was enough to ship a working assistant, but it was the wrong long-term shape for a tool that runs on your computer and is supposed to stay yours.

0.0.4 is the release that breaks that lock-in. You can stay fully local on Ollama or LM Studio, use OpenAI, Anthropic, Gemini, OpenRouter, or DeepSeek when you want a cloud model, or point Aldwin at any compatible “Other server.” Switch with /model without leaving the session.

This post is about why those choices were made, not only what landed in the changelog.

Why this release matters#

A personal assistant that only works with one vendor’s API is not really local, even if the UI sits in your terminal. Your prompts, tool results, and habits all lean on that one account. If the key expires, the region blocks you, or you want private notes to stay on loopback, you are stuck.

0.0.4 makes the model a setting, not the product. The assistant, tools, memory, skills, and connectors stay the same. Only where tokens go changes. That is the difference between “an OpenAI chat wrapper” and “an assistant you run.”

It also matches how people already work. Many already have Ollama or LM Studio running. Others already hold Anthropic or OpenRouter keys. Aldwin should meet them where they are, not force a reinstall of their model stack.

Why local-first on first run#

On launch, Aldwin probes Ollama and LM Studio before it pushes a cloud signup.

That order is deliberate. If a local model is already serving, you can chat with no account (Path A). If nothing local is found, you get a clear cloud path or install steps (Path B). If you already had an OpenAI setup from an earlier install, Aldwin keeps it and does not force you through setup again (Path C).

The goal is a quiet first run for people who care about privacy, and a short path for people who want a cloud model. Local is offered first because once you have seen “it works offline,” the cloud option is a choice, not a requirement.

Why a bundled catalogue, not an online lookup#

The first-ship providers live in a bundled providers.yaml. There is no network call to discover which vendors exist.

That keeps first run predictable offline, keeps the list reviewable in the repo, and avoids a surprise dependency on a catalogue service. The list is intentionally small: Ollama, LM Studio, OpenAI, Anthropic, Google Gemini, OpenRouter, and DeepSeek. Anything else goes through Other server or a providers: entry in config. Broadening the built-in list is a product decision, not an auto-scrape.

Why three HTTP translators and no vendor SDKs#

Aldwin talks to models through three shapes: OpenAI-compatible chat, Anthropic Messages, and Ollama’s native API. The transport is plain HTTP (httpx), not the official OpenAI or Anthropic Python packages.

That choice is about control and supply chain. Vendor SDKs pull their own dependency trees, update on their own cadence, and make “supporting another provider” look like “add another SDK.” Thin translators mean a new OpenAI-compatible host is usually a URL and a key, not a new library. The install stays smaller. agentling 0.2 is the engine dependency; the OpenAI SDK is not.

Security and packaging notes in the release follow from the same idea: fewer third-party packages in the default install, and clearer boundaries around what talks to the network.

Why environment variables win#

Cloud keys follow a simple rule: environment variables win and are never written to disk. Keys you paste during setup are checked against the provider before save, then stored under ~/.aldwin/secrets.yaml with tight permissions.

Env-wins means CI, shells, and secret managers stay authoritative. Aldwin should not silently copy a live key into a file and then ignore the environment next time. Check-then-save means a typo or revoked key fails before it becomes durable config. Friendly provider errors name the provider and what to try next, so a bad key or down host is not a stack dump.

Why Other server exists#

The built-in catalogue will never cover every lab GPU box, company gateway, or self-hosted OpenAI-compatible endpoint. Other server is the escape hatch: register a base URL, shape, and optional secret via /model or config, then treat it like any other provider.

That keeps the first-ship list honest (only vendors we actually test) without pretending the world stops at seven names.

What stays the same when you upgrade#

Upgrading from 0.0.3 keeps sessions, notes, memory, approvals, connectors, persona, and settings. No home migration runs. OpenAI users keep working. Run uv tool upgrade aldwin (or install fresh with uv tool install aldwin).

Local traffic on loopback still does not leave the machine, aside from optional web tools, remote connectors, and the light PyPI update check. Approvals, workspace rules, and the rest of the assistant behave as before. Only the model path got wider.

Limits on purpose#

Windows is still unsupported. The built-in catalogue is first-ship only. Those are conscious tradeoffs so the release could ship provider-neutral models without pretending every platform and every vendor was ready.

Try it#

Install or upgrade, run aldwin, and use /model once. Pick a local model if you have one. Pick a cloud key if you do not. The point of 0.0.4 is that both are first-class.

Full release notes: changelog v0.0.4. Getting started: Install and Models.