Central Configuration of LLM Providers
Crafting provides seamless, centrally managed integration with LLM providers at the org level. Once configured, developers can access LLMs and AI-related features from their workspaces without needing to set up credentials individually.
Configuring LLM Providers and Models
Visit Connect / LLM in the Web Console to configure LLM providers and select models.
For supported LLM providers, API keys or tokens can be stored as Secrets (recommended as Admin Only or even Not Mountable for extra security). After that, specific models can be added for use by sandboxes.
Model Purposes and Aliases
Crafting predefines a set of Purposes that allow org admins to select the right model for different kinds of tasks. Crafting agents choose models by purpose automatically.
| Purpose | Use Case |
|---|---|
CODING |
Used by agents for coding tasks. Recommend a model with strong reasoning capabilities. |
FAST |
Used for simple language processing tasks such as summarizing. |
GENERIC |
Used as the default model when the desired purpose does not have a corresponding model configured. |
For each configured model, one or more aliases can be added. Aliases allow the model to be referenced by name from agents and client applications.
Coding CLIs
For the most popular coding CLIs, they can be launched directly from a workspace without installation:
cs claude # Launch Claude CLI
cs gemini # Launch Gemini CLI
cs codex # Launch Codex CLI
If the corresponding LLM provider is configured in the org, these CLIs are automatically configured to use the provider without requiring separate credentials or login steps.
Additionally, these CLIs are pre-loaded with:
- Crafting-specific MCP tools (see MCP Servers)
- Context about the current Crafting workspace
This makes the agents aware of the Crafting environment and its specifics from the first prompt.
Workspace Auto Configuration
For supported LLM providers, enabling Auto Configure Workspace injects environment variables into all workspaces automatically. This allows LLM agents and client libraries running in the workspace to work out of the box without additional credentials or login.
See Also
- MCP Servers — centrally managed MCP servers for all sandboxes
- Agentic Sessions — LLM-driven autonomous and interactive sessions
- Secret — storing API keys securely
- Org and User — org admin responsibilities for LLM configuration