dismodels for VS Code

Chat and agent mode on shared-revenue AI models — open base models plus LoRA adapters that other people trained and published. You pay per token, the trainer whose adapter you used earns a cut, and there is no Copilot subscription and no GitHub sign-in involved.

If you train models, the same extension is where you compile a corpus into an adapter and publish it.

What you get

• dismodels models in VS Code's own model picker. Chat, agent mode, tool calling, multi-file edits and context gathering are the editor's, not a reimplementation in a side panel.
• Community adapters, priced honestly. Adapters appear as @handle/name, and the rate you see already includes the trainer's premium. There is no second number to add.
• Per-token billing, no seat. Nothing is charged for having the extension installed.
• A trainer sidebar. Your collected data, training runs and unpublished adapters, with publishing in the editor.

Requires VS Code 1.105 or newer and a dismodels account. The account is free to create; sign-in opens the page for you.

Getting started

1. Install the extension.
2. Run dismodels: Sign in from the Command Palette (Ctrl+Shift+P / Cmd+Shift+P).
3. A browser opens. Sign in or create an account, then approve the editor. You are never asked to copy or paste a token.
4. Open Chat (Ctrl+Alt+I / Cmd+Ctrl+I) and choose a dismodels model from the picker.

If the picker looks empty, you are not signed in yet — VS Code populates it silently and cannot prompt you from there. Run dismodels: Sign in and reopen it.

Using the models

Base models are listed under their usual names. Adapters are listed as @handle/name with the all-in rate in the detail line.

Agent mode works. Switch Chat to Agent and the model can read files, run the host's tools and apply edits, because this extension registers as a model provider and VS Code drives its own agent loop on top. Stop cancels the stream immediately, so a cancelled answer stops costing money.

dismodels: Manage models and account shows what this installation is holding: how many models you are entitled to, how many of them are adapters, your spend cap, and when the connection key renews.

For trainers: lessons

Lessons live in your workspace, in .dismodels/lessons/lessons.jsonl: one JSON lesson per line, in the same format as a lesson pack.

{"id": "…", "trigger": "adding a Python dependency", "guidance": "Use uv add, not pip install.", "klass": "promotable"}

Set dismodels.lessons.upload to stream to merge changed lessons a few seconds after each save instead. Set dismodels.taskKey to send X-DAM-Task on pack calls.

For trainers: the studio

Open the dismodels view in the Activity Bar. It has four sections: Lesson packs (right-click one to train a LoRA from it), plus the three below:

Start a training run from the Start a training run… item, or dismodels: Start a training run. You pick the corpus, a model name (the slug becomes @your-handle/that-name), an optional description for the model card, and the base model the adapter applies to. The run appears under Training and moves to Your models when the supplier finishes.

Publish a finished adapter from its context menu, or dismodels: Publish a model. You choose visibility — private (only you), unlisted (anyone with the slug), or public (listed on the marketplace) — and acknowledge the base model's licence. Publishing does not set a price.

Premiums, earnings and payouts are not in the editor. Use dismodels: Open trainer studio.

Commands

Settings

Pointing this at a local instance needs a URL the browser on that machine can open, because sign-in happens in the browser.

Your account and your credentials

Sign-in is OAuth 2.1 authorization code with PKCE, redirected back through the editor's own URI handler. That also works over Remote-SSH, in devcontainers, in Codespaces and in the web build, where a loopback-server sign-in would quietly fail because the browser is on a different machine from the editor.

The session appears in VS Code's Accounts menu, so signing out is where you already look for it, and other extensions can reuse the session instead of collecting their own credential.

Credentials live in the OS keychain. The extension holds a 90-day account token and uses it to mint a 24-hour inference key, renewing about half an hour before expiry. You are not asked anything after the first sign-in. Each installation registers a device id, so a renewal replaces that installation's previous key rather than leaving a trail of live ones.

Signing out revokes nothing. It forgets the credential in this editor, because signing out here is not a statement about the other places you are signed in. Revoke a device in the console.

Nothing in the editor moves money

No purchase, no plan change, no card field, no payout. Billing and payouts are links out to the console, because a card field an extension renders is a card field it is liable for, and a browser is a better place for one. Your spend is shown in the editor. It is never authorised there.

Current limitations

• No inline tab completion. Different endpoint, different latency budget. This extension is chat and agent only.
• Token counts are estimates. They are biased to over-count, so the editor drops context rather than sending a request the model will refuse.
• Images are dropped for models whose catalogue entry says they cannot take one.
• No spend in the status bar yet.

Troubleshooting

The dismodels Output channel has the detail, including key renewals and silent retries.