What Claude does in the admin panel
Claude is Anthropic’s family of language models, available through the Claude API. In Endora Commerce it has one job: interpreting instructions typed into the command palette of the admin panel. An operator writes in Polish or English, for example “set the stock of Bolts 0193 in the Default warehouse to 120”, and the model turns that sentence into calls to the tools the platform exposes to it.
The model has no access to the database or to arbitrary endpoints. It gets a closed list of tools registered in code, which currently covers:
- searching products, categories, warehouses and orders (read-only),
- setting a stock level,
- assigning products to a category and removing them from one,
- changing product status and visibility,
- changing an order status, one order or many at once.
Tools the operator has no permission for are not shown to the model at all. The list grows with the modules, because each module can contribute its own tools.
What the operator sees
After the instruction is submitted, a plan card appears: the exact operations, the current value, the number of records and a sample for bulk changes. Everything on that card is computed by the server, not written by the model. Only confirmation triggers the change, and an unconfirmed plan expires after 10 minutes. If the instruction is ambiguous, the assistant asks one clarifying question with concrete candidates instead of guessing.
Operations covering more than 50 records run in the background, and the default limit is 500 records per instruction. Every execution lands in the audit log together with the original instruction, the provider and the model name, so later on you can see that a given change was planned by Claude and by which version. A broader description of the assistant is on the features page.
Which data reaches Anthropic
The Claude API receives the text of the instruction, the descriptions of the available tools, and the results of the searches the assistant uses to identify the records named in the instruction. The API key never leaves the backend, and tool results are treated strictly as data, so a product named like an instruction will not change the assistant’s behavior.
Under Anthropic’s policy, inputs and outputs from its commercial products, the API included, are not used to train models by default. It is still worth checking the data processing terms in the agreement your organization has with Anthropic.
What to prepare on the Anthropic side
- An organization account in the Claude Console with billing set up.
- An API key created for a single workspace. A separate workspace for the admin panel lets you give it its own spend limit and keep its usage apart from other uses in the company.
- A model ID from Anthropic’s documentation, for example
claude-sonnet-5-5. All current Claude models support tool use, which is what the assistant requires.
Rate limits and the monthly spend limit belong to the organization and depend on its usage tier. New organizations start with lower limits, which rise as they build up API usage history. The details are in the rate limits documentation.
Configuration in Endora Commerce
On the credentials screen you create a configuration of type LLM, choose Anthropic (Claude) as the provider, and enter the API key and the model name. The key is encrypted, is never returned by the API after saving, and is masked on every read. You then reference that configuration in the assistant’s settings and switch the assistant on. Operators also need the prompt_actions:use permission, and every planned operation requires the same permission as its manual equivalent.
A change of model or key takes effect on the next instruction, with no restart.
What to watch for
- Model retirements. You type the model ID by hand, and Anthropic retires older models on a published schedule. After the retirement date the ID in the configuration has to be replaced.
- Direct Claude API. The integration connects to
api.anthropic.com. It does not cover Claude models offered through Amazon Bedrock or Google Cloud. - Spend limit. Once it is reached, the API rejects requests. The assistant will not build a plan, but nothing changes in your data, and the panel keeps working without the prompt mode.
When to choose Claude
What the assistant can do does not depend on the provider: the adapters for Claude, GPT and Gemini use the same tool list and the same confirmation mechanism. Claude is the natural choice when the company already has a contract and billing with Anthropic, or standardizes on these models in other tools. If you already have an account and billing with another of these providers, it is simpler to start there and change the model later in configuration. The remaining integrations are in the integrations catalog.