Creating an AI Server for Claude (Anthropic)

This guide describes how to connect Anthropic’s Claude API as an AI Server in AI Knowledge Fabric, so Claude models become selectable for Topics, GraphRAG Topics, and Agent Pipelines.

Unlike a local LM Studio or Ollama server, Claude is a cloud provider: there is no container to start, no port to configure, and an API key is mandatory.

1. Prerequisites

  1. An Anthropic account with API access — console.anthropic.com.

  2. An API key, created under Settings → API Keys in the Anthropic Console. It starts with sk-ant-api03-….

  3. Credit or an active billing plan on the account. Without it, requests are rejected with a credit balance too low error.

Handle the key like a password. It grants full access to your Anthropic account’s API usage at your expense. Do not commit it to a repository and do not paste it into a chat.

2. Create the AI Server

Navigate to Admin → AI Servers and click Add AI Server.

Fill in the form

Field Value Notes

Server Type

External (existing instance)

Required. Managed is locked to Ollama and cannot deploy a cloud API

Provider

Anthropic (Cloud)

Selecting it fills in the host and hides the port

Name

e.g. Anthropic_Claude

Must be unique. Letters and only, at most 50 characters — a space becomes

Host

api.anthropic.com

Pre-filled when you select the provider

Port

(hidden)

Cloud providers always use HTTPS on port 443. Any port value is ignored

Default Model ID

e.g. claude-sonnet-5-5

Pre-fills the model picker in the Topic wizard

API Key

sk-ant-api03-…

The key Topics, GraphRAG Topics and pipelines use to log in to Anthropic. See below

Management API Key Env Var

(leave empty)

Optional, for special cases only. See below

Click Save.

The API Key

The API Key is the only key you need. Enter it here once:

  • Topics and GraphRAG Topics get it when they start. This works the same with Docker and with Kubernetes. A changed key reaches a running topic after you stop and start it.

  • Pipelines get a copy when you pick this server for a module in the Pipeline Designer. A changed key does not reach pipelines by itself — pick the server again in each module that uses it.

  • This app uses it too, to test the connection and to load the list of models.

Management API Key Env Var (optional)

Leave this field empty, unless this app should use a different key for its own calls than the topics. Enter the name of an environment variable of this app, for example ANTHROPIC_API_KEY. This app then uses the key in that variable to test the connection and to load the list of models. Topics and pipelines still use the API Key above. If the variable does not exist, this app uses the API Key as well.

This app reads its environment variables only when it starts. A new variable therefore needs a restart of this app.

Kubernetes only: if the key should not be stored in this app at all, keep it in a Kubernetes Secret instead. See Keeping API keys out of the app.

3. Choosing a model

Enter a model ID as Default Model ID, for example claude-sonnet-5-5. After you save, the Models tab of the server lists every Claude model your account may use, with its exact ID. Take the ID from there.

Notes:

  • Use the exact ID as the Models tab shows it.

  • The Default Model ID is only a pre-selection. Each Topic and each Pipeline node can use a different model.

  • Larger models answer better but slower and cost more per question. Smaller models are faster and cheaper, and good enough for simple tasks, for example picking a search strategy.

  • Claude models can also describe pictures in documents, if a topic is set up to do so.

  • Anthropic does not provide embedding models. A Topic always needs an embedding model as well — use a local server (LM Studio, Ollama) or an OpenAI-compatible endpoint for that. See Creating an AI Server for LM Studio models.

4. Verify the connection

Open the new server’s detail page. It has three tabs:

Tab Contents

Configuration

The values you entered — editable

Models

The models the Anthropic API reports for your account

Actions

Reachability check and deletion. Start/Stop do nothing for a cloud provider

The reachability check performs an authenticated request against api.anthropic.com. Success means the host is reachable and the key was accepted.

Pull New Model is not offered — cloud models are not downloaded. That section only exists for Ollama servers.

5. Use Claude in a Topic or Pipeline

In a Topic (Topics → New Topic, model step):

  • Chat model: Anthropic_Claude → a model from its Models tab, for example claude-sonnet-5-5

  • Embedding model: a different server — Anthropic offers no embedding models

In a Pipeline (Pipeline Designer, AGENT node): pick Anthropic_Claude and a model in the model selection. This fills in provider, model, address and API key of the module. Claude models handle tool calling well, which makes them a good fit for Agents with MCP tools.

Notes:

  • Every request costs money. Unlike a local model, each chat message and each ingested document is billed against your Anthropic account. Watch usage under Admin → Metrics and in the Anthropic Console.

  • Set a timeout with a real value. Cloud requests are much faster than local inference, but long documents and deep reasoning still take time — leaving it empty is not an option.

  • The system prompt works the same for every provider. It is built from the global template and the topic’s own prompt, see Global LLM Configuration.

6. Troubleshooting

Symptom Cause and fix

Server shown as unreachable, Models tab empty

No key, or a wrong key. Check the API Key. If Management API Key Env Var is filled in, check that variable too, or empty the field

"authentication_error" / 401

Key wrong, revoked, or truncated when pasted. Create a new key in the Anthropic Console

"credit balance is too low"

No credit on the Anthropic account. Top up in the Console

Reachability check succeeds, but Topic chat fails with an auth error

The API Key field is empty, and this app used the key from Management API Key Env Var for its own check. Fill in the API Key, then stop and start the topic

A pipeline fails with an auth error after you changed the key

Pipelines keep the key they got when the server was picked. Pick the server again in each module, and save the pipeline

"not_found_error" naming the model

Invalid or retired model ID. Take the exact ID from the server’s Models tab

"rate_limit_error" / 429

Account rate limit reached. Reduce parallel Topics/Agents or request a higher tier

Document upload fails, chat works

Claude is set as the embedding model. Anthropic provides no embedding models — configure a separate server for embeddings

Related