Installation and Provider Troubleshooting
Start by identifying the failing layer. A CLI that cannot load usually indicates a local dependency problem. A CLI that starts but cannot answer usually indicates authentication, endpoint, or model-ID configuration. If ordinary replies work but tools fail, investigate model capability or protocol translation.
An error mentions usage.input_tokens
undefined is not an object (evaluating 'usage.input_tokens')A common cause is an incorrect ANTHROPIC_BASE_URL. The service returns HTML or another response that does not match the Anthropic Messages API.
The Anthropic SDK appends /v1/messages to the base URL. For example:
- MiniMax:
ANTHROPIC_BASE_URL=https://api.minimaxi.com/anthropic - OpenRouter:
ANTHROPIC_BASE_URL=https://openrouter.ai/api - Incorrect OpenRouter example:
ANTHROPIC_BASE_URL=https://openrouter.ai/api/v1, which can duplicate the/v1segment
Check the provider's current documentation, then confirm that the response is structured JSON. Do not hide an endpoint error by only increasing API_TIMEOUT_MS. See Connect a third-party model for configuration patterns.
Cannot find package 'bundle'
error: Cannot find package 'bundle' from '.../ScienceX/src/entrypoints/cli.tsx'This usually means that the Bun version does not match the repository requirement. The current repository pins bun@1.3.12 in packageManager and documents Bun 1.3.x as the supported line.
bun --versionInstall or switch to the repository-declared version, return to the repository root, and run bun install again. Avoid a generic “upgrade to latest” instruction because the latest release may be outside the currently verified range.
How do I connect OpenAI, DeepSeek, or Ollama?
ScienceX sends Anthropic Messages API requests:
- Connect directly when the provider supports
/v1/messages. - When the provider exposes only an OpenAI-compatible interface, use LiteLLM or another proxy for Anthropic → OpenAI translation.
- A local Ollama runtime usually also needs a local gateway that exposes an Anthropic-compatible interface to ScienceX.
See Connect a third-party model for the complete setup.
Can I use the Science workbench without a model provider?
Yes. Local experiment design, the simulated teaching case, data-quality analysis, and the built-in 4PL analysis do not require model credentials. Configure a provider only for AI chat and other model-dependent capabilities.
Does data stay local when I use a remote model?
Project files and run records can remain on your machine, but the prompt, context, and tool results required for a model request are sent to the configured remote endpoint. “Local-first” does not mean that every feature is fully offline. Confirm authorization and the provider's data policy before sending sensitive data.
Where are configuration and run data stored?
User-level data is stored under ~/.sciencex by default, while project-level data lives in .sciencex/ at the project root. See Configuration and data directories for credentials, shareable settings, and migration rules.