AI provider setup (BYOK)

AI naming is opt-in and requires you to bring your own API key (BYOK). The default naming mode is fully deterministic and works offline with no key. Enable AI naming when you want richer, more context-aware label names.

When to use AI naming

Default (deterministic) is the right choice for most cases:

No API key, no cost, no network call
Works in air-gapped environments
Consistent — same string always produces the same name
Validates instantly

AI naming adds value when:

You have hundreds of labels in one batch and want nuanced names
Your string values are ambiguous without surrounding context
You’re using a free local model (Ollama, LM Studio) and want richer names at $0

In either mode, every name is editable before it applies. AI naming is a suggestion, not a requirement.

Setup

Open the Command Palette (Cmd+Shift+P) → Label Alchemy: Configure Settings
Enable AI-assisted naming
Select your Provider
Select or enter a Model
Enter your API Key (skipped for Ollama, LM Studio, and unauthenticated custom endpoints)
Label Alchemy runs a capability test (sends a single test prompt). If it passes, you’re ready. If it fails, an error message shows what’s wrong.

Your API key is stored in VS Code’s encrypted secret storage — never in settings.json, never logged, never visible in plain text.

Naming scope & tuning

AI naming has two scopes (labelAlchemy.aiNamingScope):

Per-file (chunk) — Free: each file is named from its own context.
Project-wide (project) — requires a license: also reads your existing CustomLabels.labels-meta.xml, so an identical string reuses its existing label and new names follow your project’s convention. Without a license, AI naming falls back to per-file scope.

Additional settings shape the output (all only apply when AI naming is on):

Setting	Effect
`aiNamingStyle`	`concise` / `descriptive` / `domainRich` — how much component/domain context folds into a name
`labelNameCase`	`snake` (default) or `pascal` for new names (existing/reused labels are never recased)
`aiNamingGuidance`	Free-text instructions appended verbatim to the prompt (e.g. “prefix with the module name”)
`aiGenerateDescriptions`	Also write a short Description for each label (uses extra output tokens)

Whatever the model returns is always sanitized to a valid Salesforce name, and every name stays editable in the diff panel before you approve.

Supported providers

Anthropic (Claude)

Setting	Value
Provider	`Anthropic`
Recommended model	`claude-sonnet-4-6`
Fast / cheap model	`claude-haiku-4-5`
API key required	Yes
Key source	console.anthropic.com

Claude Haiku is the best price/quality tradeoff for label naming — fast, cheap, and produces clean Salesforce-style names. Sonnet if you want the best quality. Do not use Claude Opus for this task — it’s slower and more expensive with no practical naming advantage.

OpenAI

Setting	Value
Provider	`OpenAI`
Recommended model	`gpt-4o-mini`
Higher quality	`gpt-4o`, `gpt-4.1`
API key required	Yes
Key source	platform.openai.com

gpt-4o-mini is the best value for naming. gpt-4o / gpt-4.1 for more careful names on complex string sets.

Google Gemini

Setting	Value
Provider	`Gemini`
Recommended model	`gemini-2.0-flash`
API key required	Yes
Key source	aistudio.google.com

Gemini 2.0 Flash is fast and has a generous free tier. Good option if you already have a Google AI Studio key.

DeepSeek

Setting	Value
Provider	`DeepSeek`
Recommended model	`deepseek-chat`
API key required	Yes
Key source	platform.deepseek.com

Very cost-effective. deepseek-chat (V3) produces solid naming results.

OpenRouter

Setting	Value
Provider	`OpenRouter`
Model format	`<provider>/<model>` — e.g. `meta-llama/llama-3.3-70b-instruct`
API key required	Yes
Key source	openrouter.ai

OpenRouter gives you access to 100+ models from a single API key. Useful if you want to experiment or you already have OpenRouter credits. Any model that supports the OpenAI Chat Completions format works.

Ollama (local, free)

Setting	Value
Provider	`Ollama`
API key required	No
Default base URL	`http://localhost:11434`
Recommended models	`llama3.3`, `qwen2.5:14b`, `deepseek-r1:14b`
Prerequisites	ollama.ai installed and running

Nothing leaves your machine. Ollama runs the model locally. Free and private. Best choice for air-gapped setups or when privacy is a hard requirement.

Setup:

# Install Ollama, then pull a model
ollama pull llama3.3
# Ollama runs on port 11434 by default — no additional config needed

Label Alchemy connects to http://localhost:11434 automatically. Override in settings if your Ollama runs on a different port.

LM Studio (local, free)

Setting	Value
Provider	`LM Studio`
API key required	No
Default base URL	`http://localhost:1234`
Prerequisites	lmstudio.ai installed, server started

Nothing leaves your machine. LM Studio hosts any GGUF model as a local OpenAI-compatible server. Same privacy guarantees as Ollama.

Setup:

Download and install LM Studio
Download a model (Models tab)
Go to Local Server → Start Server (default port 1234)
Set Provider to LM Studio in Label Alchemy settings

Custom OpenAI-compatible endpoint

Setting	Value
Provider	`Custom`
Base URL	Your endpoint URL
API key	Optional (depends on your endpoint)
Model	Whatever your endpoint accepts

Use this for:

A self-hosted vLLM or LiteLLM proxy
A corporate proxy in front of a supported model
Any other service that speaks the OpenAI Chat Completions API

Model requirements

Label Alchemy sends a single chat message to the model and parses the response as a raw string (not JSON). Any model that can follow instructions well enough to return a short underscore-separated name is compatible.

Minimum viable models: ~7B parameters or equivalent quality. Models below 7B tend to produce inconsistent or non-compliant names.

Capability test

When you configure a provider, Label Alchemy runs a quick test: it sends the message Return only this text: Hello and checks whether the response contains Hello. If the test fails:

Wrong base URL — check that the URL matches your provider’s endpoint exactly
Invalid API key — the key may have been copied with whitespace or is expired
Model not available — the model name may be wrong or not loaded (Ollama/LM Studio)
Server not running — start Ollama or LM Studio server first

See Troubleshooting for more on capability-test failures.

Privacy

When you use an API-based provider (Anthropic, OpenAI, etc.), the string values from your source files are sent to that provider’s API. This is necessary for the model to suggest names.

If your Salesforce code contains confidential business strings, use Ollama or LM Studio — both run locally and nothing leaves your machine.

Keys are stored in VS Code’s encrypted secret storage (vscode.SecretStorage). They are never written to disk in plain text and are not accessible to other extensions.

See the full privacy policy for details on what is and isn’t sent.