The Anthropic Claude service provides advanced AI capabilities for natural language processing and understanding, enabling seamless integration into applications for enhanced user interactions.
Changelog
10 versions
v1.2.0
Latest
- SendPrompt, TransformTextToJSON, AIAgent: added `usage` (input/output/cache tokens), `stop_reason` and `model` to the output. AIAgent sums `usage` across all API calls of the tool-use loop.
- TransformTextToJSON: fixed a `ReferenceError` on the required Output JSON Schema check; when the response is cut off by `max_tokens`, the component now outputs `stop_reason: max_tokens` instead of failing.
- AIAgent: any component connected to the `tools` port is a tool the agent calls directly, with no `ToolStart`/`ToolOutput` around it; fields set to the new `Model Defined Parameter` variable are filled by the model, a field set to `Prompt` gets the prompt, other values are sent as configured. Components connected behind it run as further steps of the same tool. MCP servers can be connected to the port too. A `ToolStart` on a port with other tools is called through its webhook URL, so the other tools do not run; a port with `ToolStart` chains only works as before.
- AIAgent: all tool calls of a turn are answered (the agent handled only the first one, which made the Claude API reject the next request), and the tool-use loop stops after 20 turns. Calls of different tools run at the same time. A `ToolStart` chain now has 120 seconds to answer instead of 60 (`TOOLS_OUTPUT_POLL_TIMEOUT`); a timeout is reported to the model as an error text. The tool names the model sees for `ToolStart` chains changed to include the component label.
- AIAgent: new Max Tokens input. Without it the agent now requests 8192 tokens instead of 1024 (4096 for the Claude 3 models, their limit): models with extended thinking used the whole 1024 tokens to think and answered with an empty text. Longer answers cost more output tokens.
- AIAgent: a `ToolOutput` that comes after its tool call has timed out is removed from the flow state with the next message of the agent, instead of staying there.
v1.1.4
- MakeApiCall: Request Headers and Query Parameters set in the Designer are now sent. They were silently dropped: the key-value field stores its rows as a JSON object string, which the component did not parse.
v1.1.3
- Output examples for SendPrompt.
v1.1.2
- AIAgent: removed the unused icon.svg duplicate (the icon comes from component.json); added output examples.
v1.1.1
- Added a missing output example to ListModels.
v1.1.0
- Added MakeApiCall component for performing arbitrary authorized API calls to the Anthropic (Claude) API.
v1.0.5
- Updated langchain npm dependencies.
v1.0.4
- Added validation for required fields in components.
v1.0.3
- Fixed an error `JSON schema is invalid` that occurred when the AiAgent used a tool with an empty parameters object.
v1.0.2
- Initial version
Authentication
| Type | Details |
|---|---|
| API Key |
API Key (text)
Log into your service account and find API Keys page.
|
Actions
5
A I Agent
Build an AI agent responding with contextual answers or performing contextual actions. Connect a component to the 'tools' output port to let the agent call it as a tool, and mark the fields the model should fill with the 'Model Defined Parameter' variable; components connected behind it run as further steps of the same tool. For a tool with its own description and logic (branches, loops), connect a 'Tool Start' to the 'tools' port and end every branch with 'Tool Output'. An MCP server on the 'tools' port gives the agent all of its tools.
